1.1 How large language models actually work — in plain English

You have probably typed a question into one of these tools, received a confident answer, and then not been quite sure how much to trust it. That uncertainty is not a failing on your part. Nobody explained what the thing is actually doing, so you have been guessing. Five minutes on the mechanism prevents most of the bad habits before they form.

What happens when you press enter

ChatGPT, Claude, Gemini and Perplexity are all built on large language models. A large language model is a piece of software that has read an enormous amount of text and learned, from the patterns in that text, which words tend to follow which. When you type something, it continues your text in the most likely way, one small piece at a time, until it has produced an answer.

That is the whole trick. It is not looking your question up in a database. It is not checking a fact against a source unless it has been given one or connected to one. It is producing the words that most plausibly come next, given everything it has learned and everything it can currently see.

What it can see: the context window

"Everything it can currently see" has a name: the context window. Think of it as the desk the model is working at. On the desk are your prompt (the text you type), anything you paste or upload, and the conversation so far. Nothing else. Not your inbox, not your shared drive, not your company's style guide, not this morning's news, unless you put it there or your organisation has connected the tool to those sources.

The desk is large but not infinite. In a very long conversation, details from the beginning lose their influence, which is one reason a chat that has been running for an hour starts to drift. Starting a fresh chat with a clean brief is often quicker than arguing with a tired one.

Why it sounds right when it is wrong

The model was trained to produce text that reads well. Reading well and being true are two different properties, and only the first was the target. So when the model lacks a fact, it does not go quiet or hesitate. It produces the most plausible-sounding fact, in the same even tone it uses when it is correct. Fluency is not evidence of anything. It is the one thing the machine was built to deliver.

Priya, a marketing manager at Brightwater Property in Brisbane, asked a tool to "write our standard welcome email for new tenants". She got a polished email in seconds. It mentioned an out-of-hours maintenance line Brightwater does not have and a welcome pack they stopped sending years ago. The tool had never seen a Brightwater email. It wrote the welcome email a plausible property company would send, because that was all the desk contained: her one sentence and a great deal of general pattern.

Three consequences you will use every day

• Give it material. If the answer depends on something the model could only know by being told (your policy, your client, your figures, your previous draft), put it on the desk. Paste it, upload it or describe it. Day 2 is largely about doing this well.
• Be specific. A vague prompt leaves the model to fill the gaps from general pattern, and general pattern is where invented maintenance lines come from. Say who the audience is, what you want, how long it should be and in what tone.
• Check. Anything factual that the model produced from memory rather than from your material needs a second look before it leaves your desk. Lesson 1.2 shows you how to spot which parts those are.

I am going to give you a task. Before you do it, answer three questions as a short list:
1. What information would you need from me to do this well that I have not given you?
2. Which parts of your answer would come from material I provide, and which from your general knowledge?
3. What is the most likely way your answer could be wrong or out of date?
Then stop and wait for my reply before doing the task.
The task: [DESCRIBE THE TASK IN ONE OR TWO SENTENCES].

1.2 What AI is brilliant at, what it fakes, and how to spot a hallucination

You have probably had both experiences already. One afternoon the tool turns a rambling voice note into a tidy project update in thirty seconds. The next morning it invents a supplier, a clause number and a percentage, all in one confident paragraph. It is the same machine doing the same thing in both cases: producing the most plausible next words. The only difference is whether "plausible" happened to coincide with "true".

Knowing in advance which kind of task you are handing it is the most useful habit in this course.

Where it is genuinely strong

The model is at its best when the raw material is already on the desk and the job is to work on it.

• Transforming text you give it — summarising, shortening, re-ordering, changing the tone, translating "engineer" into "client" and back.
• Drafting from a clear brief — first versions of emails, agendas, job adverts, FAQs, proposal outlines. First versions, not final ones.
• Thinking out loud with you — listing options, arguing the other side, suggesting the questions you have not asked, predicting what a sceptical reader will object to.
• Explaining — a contract clause, a spreadsheet formula, a policy document, a technical email, at whatever level you ask for.
• Structuring — turning a mess of notes into headings, a comparison into a table, a decision into a list of criteria.

Where it fakes it

Anything that depends on a fact it was not given: names, numbers, dates, prices, citations, case law, what happened last month, what your own company's policy says, who said what in a meeting it did not attend. The industry word for a confident, fluent statement that is simply made up is a hallucination. The word is slightly misleading because it suggests a malfunction. It is not a malfunction. It is the normal mechanism from 1.1 running without the facts it needed.

Tom, an operations manager at Northwind Logistics in Leeds, asked for "the current rules on driver rest breaks" and received a clean, numbered list. Two of the numbers were wrong. Nothing in the answer looked wrong, and that is the point: a hallucination does not come with a warning light. The fix was not a cleverer prompt. It was pasting the actual regulation text and asking the tool to summarise that instead.

The tell-tale signs

You cannot verify everything, so learn what tends to be invented. Treat each of these as a flag that means "check this first".

• Over-specific detail you did not provide. You asked for a draft and it supplied a street address, an invoice number or a colleague's job title. Where did that come from?
• Confident citations. A named report, an author, a journal, a section of an Act, a web address. These are among the most frequently fabricated things a model produces, and they look the most authoritative.
• Suspiciously round numbers. Real figures are untidy. A model filling a gap reaches for a tidy one.
• "As of" dates. Phrases such as "as of" followed by a year, or "the current rate is", signal that the model is reaching for memory rather than reading your material, and that memory has a cut-off date.
• Agreement with your framing. If you ask "why is option B better?", it will tell you why option B is better. Ask it to argue the other side before you decide anything.

Here is a draft you wrote for me: [PASTE THE DRAFT].
Go through it and list every specific claim (names, numbers, dates, prices, quotations, citations, legal or policy statements) as a checklist. Mark each one as either "from the material I gave you" or "from your general knowledge". For anything marked general knowledge, add one line on how I could verify it. Do not rewrite the draft and do not add new claims.

1.3 The confidentiality rules: what you can and can't paste into a public tool

The quickest way to lose permission to use AI at work is one careless paste: a client's name and fee, attached to a request for "a friendlier version of this email". The box looks private, so it feels harmless. This lesson gives you a sorting habit that takes under a minute.

Four categories

• Public — already published: your website, press releases, public reports. Paste freely.
• Internal — not secret, not published: meeting notes, process documents, draft plans. Fine in an approved tool; in a consumer tool, strip anything that identifies people.
• Confidential — client details, commercial terms, unreleased figures, anything under a non-disclosure agreement, HR matters, security information. Not for consumer tools as it stands.
• Personal data about identifiable people — a name or other identifier attached to anything else: salary, health, performance, a complaint, contact details. Privacy law is built around this category; it covers colleagues as well as customers.

The stranger test

Not sure which category? Ask: would I email this, as it is, to a competent stranger at another company? If not, do not paste it into a consumer AI tool as it is. The test works because it removes the false privacy of a chat box: you are sending text to another company's servers.

Anonymise first, then paste

The model almost never needs the real values; it needs the shape. A fee-increase letter reads the same whether the client is "Harbourline Dental" or "[CLIENT]". So replace names with roles, companies with [COMPANY A] and [COMPANY B], amounts and identifiers with placeholders, and dates with month and year. Keep a key so you can restore the real values afterwards. Do the first pass yourself with find-and-replace; use the prompt below for a second pass only in a tool your employer has approved for that data.

Mei, a senior accountant at Maple Ridge Accounting in Vancouver, needed to explain a complicated tax position to an anxious client. Instead of pasting the file, she wrote a six-line summary with [CLIENT], [THE PROPERTY] and [AMOUNT] in place of the specifics. The tool never saw a name or a dollar figure, and the explanation was no worse for it.

Consumer tools versus business tiers

In general terms: free and personal versions may, depending on your settings, use your conversations to improve future models; most offer a way to opt out and a way to chat without the conversation being kept. Business and enterprise tiers, the versions sold to organisations, typically commit contractually not to train on your data and give administrators control over what staff can connect. The details change often, so read your tool's current terms.

The law, in general terms

United Kingdom. UK GDPR governs personal data about identifiable people: use it for a lawful reason and the purpose it was collected for, use as little as possible, and keep it secure. Pasting it into a third-party tool is a form of sharing, so your organisation must have agreed to that first.

Canada. PIPEDA is the federal private-sector law, built on consent, reasonable purposes and safeguards; some provinces have their own laws that apply instead of or alongside it. Do not disclose personal information to a new party, including an AI provider, for a purpose the person would not reasonably expect.

Australia. The Privacy Act 1988 sets out the Australian Privacy Principles on collecting, using, disclosing and securing personal information. It covers many organisations but not all, so check whether yours is one. Sending personal information overseas carries extra obligations, and most AI tools are overseas providers.

New Zealand. The Privacy Act 2020 applies to almost every organisation and sets out principles on collection, use, disclosure, security and sending information overseas. It also requires serious privacy breaches to be reported, and a paste you cannot take back could be exactly that.

None of this is legal advice; it is general guidance from course writers, not lawyers. Your privacy officer or legal team has the final word, and your employer's policy comes first.

Anonymise the document below so I can work on it safely elsewhere. Replace every person's name with a role in square brackets, such as [CLIENT] or [FINANCE DIRECTOR]. Replace every organisation name, except well-known public brands, with [COMPANY A], [COMPANY B] and so on, using the same placeholder for the same organisation throughout. Replace every monetary amount with [AMOUNT], and every account number, invoice number, reference, address, email address and phone number with [REF]. Reduce dates to month and year. Keep the structure, headings, tone and the meaning of every sentence intact so the document is still useful for drafting and analysis. Output the anonymised document first, then a short key listing each placeholder and the type of thing it replaced, without the original values.
[PASTE THE DOCUMENT]

1.4 Set up your toolkit in ten minutes

Everything from tomorrow onwards assumes you have accounts open, somewhere to keep what works, and a short set of rules you can actually remember at five o'clock on a Thursday. People who skip this step spend the week re-typing prompts they have lost and hesitating over what they are allowed to paste. Ten minutes now removes both problems.

Step 1: four free accounts

Create free accounts for ChatGPT, Claude, Gemini and Perplexity (an AI answer engine that searches the web as it replies). The free tiers are enough for the whole course. You are setting up four rather than one because they have different strengths, and by the end of the week you will know which you reach for first for which kind of task. The cheat-sheet in Bonus 3 summarises it for you.

Use your work email only if your employer allows it, and do not connect any tool to your work files, calendar or email yet. That decision belongs after you have read your policy (1.3), not before. Higgsfield, Canva, Notion AI and Zapier come later in the week; leave them for now.

Step 2: a scratchpad

Open a new document in whatever you already use (Word, Google Docs, Notion, OneNote, a plain text file) and call it something like "AI scratchpad". Give it four headings: Prompts that worked, Results worth keeping, Things it got wrong, Questions for later. Date each entry. Whenever a prompt produces something you would actually use, copy the prompt, not just the result, under the first heading.

Aroha, a consultant at Kōwhai Consulting in Wellington, spent her first fortnight with AI typing one-line questions into whichever tool was open, then losing the good answers in a scroll of chat history. The scratchpad changed that. Within a week she had a page of prompts she reused daily, and her worst mistakes were written down where she could see the pattern. By the end of this course yours will be the start of a personal prompt library, alongside the ready-made prompts in Bonus 1.

Step 3: your three-rules card

At the top of the scratchpad, or on a card beside your screen, write three rules you will follow without thinking. Three, because you will remember three, and you will not re-read a policy document under deadline. Most people's cards start with something like:

• Nothing confidential, and no personal data, in consumer tools.
• Verify every fact before it leaves my desk.
• One task per prompt.

Adapt the wording to your own work. If your role carries a specific risk (client names in a law firm, patient details in a clinic, unreleased figures in finance), make that your first rule, in your own words.

How to use this course

• Twenty minutes a day, at the same time if you can. Each day is designed to fit before the first meeting or in the lull after lunch.
• Apply each lesson to real work the same day. The "Do this now" tasks are the course. The reading is preparation for them.
• Fill in the worksheet. It saves in your browser. Writing the answer down is what turns a lesson into a habit.
• If you miss a day, do not double up. Pick up where you left off, and do not try to read all seven days today.

I am a [YOUR ROLE] at a [TYPE OF ORGANISATION, NO NAME] in [COUNTRY]. I am learning to use AI tools properly at work. In no more than 150 words, tell me:
1. The three kinds of task in my role you are most likely to help with well.
2. The two kinds where I should be most careful about trusting your answers.
3. One question you would ask me before helping with my next task.
Be direct, avoid flattery, and keep each point to one line.

2.1 The CRAFT framework: Context, Role, Action, Format, Tone

You would never walk up to a new colleague, say "write the update" and walk off. You would tell them what the project is, who the update is for, what has changed, how long it should be and whether the mood is reassuring or urgent. Then you would expect a decent first draft. AI tools need the same briefing, and most of the frustration people feel with them comes from skipping it.

CRAFT is a checklist for that briefing: Context, Role, Action, Format, Tone. You will not need all five every time. But when an answer disappoints, one of the five is almost always missing, and knowing which one tells you what to add.

The five elements, and what happens without them

• Context — who you are, what has happened, who will read the result and what they already know. Without it, the tool fills the gap with the most average version of your situation, so you get text that could be about any project in any company.
• Role — the hat you want it to wear: "an experienced project manager", "a plain-English editor". Without it, you get the default general assistant, which is cautious, wordy and bland. A role sets the level of expertise and the kind of judgement you want applied.
• Action — the specific thing you want done, starting with a verb: draft, summarise, list, compare, rewrite. One action per prompt where you can manage it. Without it, "help me with this" gets interpreted in the broadest possible way.
• Format — the shape of the output: length, structure, bullets or paragraphs, whether you want a subject line. Without it, you get the tool's house style, which is usually longer than you want and sprinkled with headings you did not ask for.
• Tone — how it should sound: calm, direct, warm, formal, no jargon. Without it, you get the default voice: polished, upbeat and fond of words like "exciting" and "robust".

A weak prompt and its rewrite

Priya leads an office fit-out project at Brightwater Property in Brisbane. The electrical contractor has slipped by two weeks and she needs to tell her director. Her first prompt was: "Write an email updating my boss on the office fit-out project."

She got three paragraphs of generic reassurance, an invented timeline and a sign-off that read "Warm regards, [Your Name]". Fixing it took longer than writing from scratch would have. Here is the same request as a CRAFT brief.

Context: I lead the office fit-out project at Brightwater Property in Brisbane. The electrical contractor has pushed second-fix work back by two weeks because of a parts delay, so the move-in date moves from 14 to 28 November. The budget is unchanged. My director, Helen, knows the project well and dislikes long emails.
Role: Act as an experienced project manager who writes calm, clear status updates.
Action: Draft an email to Helen covering the delay, its cause, the new date, what I have already done about it, and the one decision I need from her: whether to tell tenants now or after the revised schedule is confirmed on Friday.
Format: Under 150 words. A subject line, then three short paragraphs. No bullet points, no headings.
Tone: Direct and matter-of-fact. One sentence of apology at most, no reassurance padding.
Before you answer, list any information you are missing.

The reply asked one question first: had the tenants already been given the original date? Priya answered in a line, and the email that followed went out with two words changed. Notice that the brief is longer than the email it produced. That is normal, and it is still faster.

The reusable template

Keep this somewhere you can paste from. Five labelled lines, then the optional closer. You can drop Role when the task is simple, and drop Tone once the voice file from lesson 2.4 covers it.

Context: [WHO YOU ARE, WHAT HAS HAPPENED, WHO WILL READ THE OUTPUT AND WHAT THEY ALREADY KNOW].
Role: Act as [THE EXPERT OR COLLEAGUE YOU WANT], someone who [ONE QUALITY YOU VALUE, SUCH AS WRITING IN PLAIN ENGLISH].
Action: [DRAFT / SUMMARISE / LIST / COMPARE / REWRITE] [THE THING YOU NEED], making sure it covers [WHAT MUST BE INCLUDED].
Format: [LENGTH LIMIT]. [STRUCTURE, SUCH AS A SUBJECT LINE AND THREE SHORT PARAGRAPHS]. Leave out [WHAT YOU DO NOT WANT].
Tone: [THREE WORDS, SUCH AS CALM, DIRECT, WARM]. The reader is [BUSY / SCEPTICAL / NEW TO THE TOPIC].
Before you answer, list any information you are missing.

2.2 Why your first prompt is never the final one: iterating in three moves

You read the first draft and it is fine. Not good, fine. At this point most people do one of two things: accept it, and send something that reads like everyone else's; or delete the lot and start again with a longer prompt. Both waste the most useful property of these tools, which is that they can revise what they just wrote, instantly, for as long as you have the patience.

Treat the first answer the way you would treat a colleague's first draft: the start of a short conversation. Three moves cover almost every situation, and each is one short message. You usually need two or three turns, not ten. If you are on turn six, the brief was the problem, so go back to CRAFT.

Move 1: critique

Ask: "What is weak about this and why?" It feels odd to ask the tool to mark its own work, but it is far better at judging a draft against a standard than it is at producing the ideal draft blind. Give it the reader's lens: "as the busiest person on the recipient list would read it". You will get a list of padding, vague claims and buried asks. Do not accept the whole list. Pick the two points that matter and say "fix those two". That keeps the judgement with you.

Move 2: constrain

Add one rule or limit. One, not five. "Under 120 words." "The request in the first sentence." "No adjectives." "Do not use the word 'excited'." A single constraint forces a genuine rewrite; several at once produce a muddle in which each is half honoured. If the result still overshoots, add the next constraint in the next message.

Move 3: compare

Ask for three variants that differ in a named way: "three versions that open differently, one leading with the deadline, one with the benefit to the reader, one with a question". The named difference is the whole trick. Ask for "three versions" with nothing else and you get three near-identical paragraphs. Then pick one, or merge: "the opening of 2 with the body of 1". You now have something you chose, not something you were handed.

Worked example

Callum, a partner at Maple Ridge Accounting in Calgary, needed the opening paragraph of a client newsletter about the busy season. The first draft began: "As tax season approaches, we at Maple Ridge Accounting are excited to share some important reminders to help you stay on top of your filing obligations."

Critique: the tool pointed out that the paragraph opened with the firm rather than the client, that "excited" rang false next to deadlines, that "important reminders" promised nothing specific, and that no sentence told the reader what to do. Callum asked it to fix the first and last points.

Constrain: "Under 50 words, and the client's action in the first sentence."

Compare: three openings, one leading with the date, one with the consequence of lateness, one with a question. He took the date-led opening and the second version's closing line. The result: "Send us your tax slips by the end of March and your return will be filed with time to spare. Late paperwork means a rushed return and, sometimes, a penalty. Reply to this email and we will send you the checklist." Four messages, a few minutes, and it sounds like a firm that knows its clients.

Read the draft you just wrote as [THE BUSIEST PERSON WHO WILL RECEIVE IT] would read it. List what is weak about it and why: padding, vague claims, anything that sounds generic, and whether the main point or request arrives early enough. Be blunt and specific, and quote the words you would cut. Do not rewrite yet. I will tell you which points matter and then ask you to fix those.

Rewrite the draft with one new rule: [THE RULE, SUCH AS "UNDER 120 WORDS" OR "THE REQUEST IN THE FIRST SENTENCE" OR "NO ADJECTIVES"]. Keep everything that was already working. Do not add new content to fill any space the rule frees up. At the end, tell me in one line what you had to cut or change to meet the rule.

Give me three versions of this [PARAGRAPH / EMAIL / POST] that differ only in [THE NAMED DIFFERENCE, SUCH AS HOW THEY OPEN: ONE WITH THE DEADLINE, ONE WITH THE BENEFIT TO THE READER, ONE WITH A QUESTION]. Keep the length and the facts identical across all three. Label them 1, 2 and 3, and under each add one line on who it would suit best. I will then pick one or ask you to merge parts of two.

2.3 Show, don't tell: examples and reference material

You have probably tried describing what you want. "Make it punchier. More like our usual style. Not too formal, but not casual either." And you got something that was none of those things, because those words mean something different to every reader, including the artificial one. Describing a style is hard for people too; it is why brand guidelines are so long and so widely ignored.

What works is showing. These tools are very good at matching a pattern they can see and poor at guessing one you are holding in your head. Three techniques cover most of what you need.

1. Paste an exemplar and ask it to match

Find a piece you or a colleague wrote that has exactly the shape and feel you want: last month's report, a proposal that landed well, an email that got a quick yes. Paste it and ask the tool to match its structure, its style, or both, using the new content you supply. Say which you mean. "Same structure" gets you the section order and proportions; "same style" gets you the sentence rhythm and vocabulary. Ask for both when you want both.

2. Supply the source material and fence it off

When the output must be factually right, paste the documents it should draw on and add one instruction: "Use only the material provided. If something is missing, say so." Without this line the tool will quietly top up from general knowledge, and you will not be able to see where your facts end and its guesses begin. With it, gaps become visible as gaps, which is exactly what you want.

3. Give two or three examples to teach a pattern

For repetitive transformations, such as turning raw meeting notes into action lines in your team's format, do not describe the format. Give two or three real before-and-after pairs, then the new input. The tool works out the pattern from the examples, including the things you would never think to mention. Two or three is usually plenty; what matters is that the examples are consistent with each other, because inconsistencies get copied just as faithfully as the good habits.

Why this beats a long description

A description is your interpretation of a pattern. An example is the pattern. It carries the length, the rhythm, the order and the small habits all at once, and you can check the result against it side by side. It is also quicker: the examples are already sitting in your sent folder.

Worked example

Aroha, a consultant at Kōwhai Consulting in Wellington, writes a one-page engagement summary at the end of every client project. For a new one she pasted a summary she was proud of as the exemplar, then the raw material: the original proposal, her notes from the final presentation and the client's feedback email, with the "use only the material provided" line.

The draft came back in her four-section structure, kept her habit of ending with one line on what she would do differently, and flagged that nothing in the material stated the final fee, so it left a marked gap instead of inventing a figure. That one flag saved her an awkward correction later.

Below are two things. First, an EXAMPLE of a finished [REPORT / EMAIL / SUMMARY] whose structure and style I want you to match: the same section order, similar paragraph lengths, the same level of formality, and the same kind of opening and closing. Do not reuse its content.
Second, the SOURCE MATERIAL for the new one.
Write the new [REPORT / EMAIL / SUMMARY] using only the source material. Where the structure needs something the source does not contain, write [MISSING: WHAT IS NEEDED] in its place rather than guessing. Finish with a short list of the gaps you found.

EXAMPLE:
[PASTE THE EXAMPLE]

SOURCE MATERIAL:
[PASTE THE NOTES, DOCUMENTS OR EMAILS]

I am going to show you three examples of how I turn [THE RAW INPUT, SUCH AS MEETING NOTES] into [THE OUTPUT, SUCH AS ACTION LINES]. Study the pattern: what I keep, what I drop, how I phrase things and how long each one is. Then apply the same pattern to the new input at the end. Do not explain the pattern or add commentary; produce only the output, in the same format.

EXAMPLE 1
Input: [PASTE]
Output: [PASTE]

EXAMPLE 2
Input: [PASTE]
Output: [PASTE]

EXAMPLE 3
Input: [PASTE]
Output: [PASTE]

NEW INPUT:
[PASTE]

2.4 Build your personal voice file

You can tell when a colleague's email was written by AI, and they can tell when yours was. Everything the tool writes sounds like the tool until you tell it otherwise. You can fix that each time with the Tone line in CRAFT, but you will describe your voice slightly differently every time and get slightly different results.

A voice file is the one-off fix. It is a short document, half a page to a page, that describes how you actually write. You create it once, keep it where the tool can see it, and say "write this in my voice".

What goes in it

• Samples — two or three short pieces of your real writing, to different readers, with names and confidential detail removed.
• Sentence and paragraph length — short and brisk, or longer and measured; one idea per paragraph or several.
• Preferred phrases — the words you genuinely use: "Quick one", "Happy to talk this through", "Over to you".
• Banned phrases — what you would never write: "reach out", "circle back", "I hope this email finds you well", "delighted to announce".
• Greetings and sign-offs — by relationship: "Hi Tom" and "Thanks" for colleagues, "Dear Mrs Okafor" and "Kind regards" for clients.
• Formality by audience — how you shift between your team, your manager, clients, suppliers and the public.
• Mechanics — British spelling, contractions or not, how you write dates and numbers, whether you ever use bullet points in an email.

How to create it

Collect three to five real samples, written by you without AI help, to at least two different kinds of reader. Strip names and anything sensitive. Paste them into the prompt below. Then read the result sceptically: it will probably call you "warm and professional", which describes everyone. Push back: "Which of these observations could apply to anyone? Replace them with things that are only true of these samples." Finally, add the banned list by hand, because the tool cannot see what you avoid.

Fatima, practice manager at Harbourline Dental in Plymouth, writes to patients, suppliers and the practice owner, and sounds different to each. Her voice file has three short audience sections. The line that changed the most output was one she added herself: "Never open a patient email with an apology; open with what happens next." The tool had been apologising in every single one.

Where to keep it

The simplest option is a note you paste at the start of a chat. The better option is the feature in each tool that lets you save standing instructions so every new chat starts with them: Projects in Claude, custom instructions or a custom GPT in ChatGPT, the equivalent in Gemini. Save the voice file there and every draft in that space will use it without being asked.

One caution: saved instructions apply to everything in that space. Keep your writing space separate from, say, a space for analysing spreadsheets, or your data summaries will start sounding like patient letters. Re-read the file every few months, and add a banned phrase whenever you catch the tool using one.

How to use it

If it is saved as standing instructions, just write your CRAFT brief; the voice file replaces most of the Tone line. If not, start the message with "Write this in my voice:", paste the file, then add the brief. Either way, still read before you send. The file gets the draft close to you; it does not make it you, and your readers know your voice better than any document does.

Below are [THREE TO FIVE] pieces of writing I wrote myself, each for a different reader. Analyse them and write a "voice file": a short guide another writer could follow to sound like me. Cover:
1. Typical sentence and paragraph length, with one quoted example of each.
2. Words and phrases I use often, quoted from the samples.
3. How I open and close, for each type of reader you can identify.
4. How my formality shifts between readers, with examples.
5. Habits of spelling, punctuation and formatting: contractions, dashes, lists, dates, numbers.
6. Three things I never seem to do.
Every observation must point to evidence in the samples. Do not use words like "professional", "warm" or "clear" unless you quote what makes it so. Keep the whole file under 400 words.

SAMPLE 1, written for [WHO IT WAS FOR]:
[PASTE]
SAMPLE 2, written for [WHO IT WAS FOR]:
[PASTE]
SAMPLE 3, written for [WHO IT WAS FOR]:
[PASTE]

2.5 Common prompting mistakes and the fixes

You now have the pieces: a brief, three moves, examples, a voice. This lesson is about how those pieces fall apart under time pressure, because they will. Most disappointing output traces back to one of six habits, and each has a one-line fix you can apply before you press enter.

The six mistakes

• Vague asks. "Can you help me with this report?" Fix: name the verb and the object: "Summarise this report's recommendations in five bullets for someone who has not read it."
• No format specified. You get eight paragraphs when you wanted three lines. Fix: always state a length and a shape, even roughly: "under 150 words, two paragraphs, no headings".
• One giant task instead of steps. "Research the market, write the proposal and draft the slides." Fix: one task per message, and check each result before asking for the next.
• Overlong prompts that bury the ask. Three paragraphs of background with the actual question somewhere in the middle. Fix: put the ask in the first line or the last, and cut any context that would not change the answer.
• Accepting the first answer. It was fine, so it went out. Fix: one critique round, minimum, before anything leaves your desk.
• Forgetting who will read the output. A note to your team and a note to a client are different jobs. Fix: name the reader and what they already know in every brief.

Before and after

Jordan coordinates operations at a Northwind Logistics depot in Leeds. From Monday, loading bays will be allocated through a booking system instead of first come, first served, after a near-miss last month. Jordan's first prompt, typed between phone calls, read:

"Can you help me write something about the new loading procedure? Basically from Monday the bays are going to be allocated differently and there was an incident last month so we need everyone to follow it, management is quite keen on this, can you also do a poster and a checklist and maybe a short training outline, thanks."

Count the faults: a vague ask ("write something"), no format, four tasks in one, the actual change buried mid-sentence, and no mention of who will read it. The rewrite below took two minutes and fixed all five. The sixth habit, accepting the first answer, was dealt with by a critique round afterwards.

Context: I coordinate operations at a Northwind Logistics depot in Leeds. From Monday, loading bays are allocated through a booking system rather than first come, first served, following a near-miss last month. The readers are forty drivers who read messages on their phones between jobs.
Action: Draft a notice telling them what changes, exactly what to do differently from Monday, and who to ask if a booking fails.
Format: Under 120 words. Plain sentences, the change in the first line, no headings.
Tone: Direct and respectful. No corporate language and no mention of management being keen.
I will ask for the checklist in a separate message once this is right.

Notice that Jordan dropped the Role line. That is fine; the brief was simple enough without it. CRAFT is a checklist, not a form you must fill in completely.

A checker you can run first

When you are in a hurry, let the tool catch the faults for you. Paste your brief into the prompt below before you run it. It cannot check whether you will accept the first answer; that one is on you.

Before you do anything else, check the brief below against five common faults and tell me which apply: (1) the ask is vague or has no clear verb, (2) no length or format is given, (3) it contains more than one task, (4) the main request is buried under background, (5) it does not say who will read the output or what they already know. For each fault you find, suggest the one line I should add or change. Do not carry out the brief yet; wait for my corrected version.

BRIEF:
[PASTE YOUR PROMPT]

3.1 Emails in under five minutes

You know which emails take the longest. Not the forty routine replies, but the three you keep reopening: the one saying no, the one chasing something for the third time, the one carrying news the other person will not like. The words are not the hard part. The hard part is deciding what you actually want to say and how it will land, and that is exactly where a drafting tool earns its keep.

Start with three bullet points

Before you ask for anything, write three bullets: who the email is to and your relationship with them, what you need them to do or know, and anything they need in order to do it. That is the Context and Action from your CRAFT brief, and it takes less than a minute. Then ask for a draft in your voice file's tone, at the length the situation deserves.

Priya, an HR adviser at Harbourline Dental in Brisbane, had to tell a locum dentist that her start date had moved back a fortnight because the surgery fit-out was late. Her bullets were: "Locum, first time working with us, starts Monday week. Start moves from the 3rd to the 17th. Fit-out delay, nothing to do with her. Confirm she can still do the 17th; offer a call." The first draft needed two edits and went out a few minutes after she opened the chat.

Shift the tone, not the brief

Once you have a draft, do not re-brief. Ask for the same email warmer, firmer or shorter, and say what must not change: "Same email, firmer, keep every fact and the Friday date." If you cannot decide, ask for two versions side by side. Three adjustments cover almost everything you will ever need: warmth, firmness and length.

The three difficult emails

• Declining — say no in the first two lines, give one honest reason or none, and offer whatever you genuinely can. Do not let the tool pad it with apology.
• Chasing — assume good faith, restate the ask in one sentence, give a specific date, and make the reply as easy as possible (a yes, or a date).
• Delivering bad news — lead with the news, own your part of it, say what happens next and when they will hear from you again. Never bury it under a paragraph of context.

You are helping me write a polite but firm chasing email.
Context: I asked [NAME, THEIR ROLE] for [WHAT YOU ASKED FOR] on [DATE OF FIRST REQUEST] and followed up on [DATE OF LAST FOLLOW-UP]. I have had no reply. I need it by [DEADLINE] because [ONE-LINE CONSEQUENCE].
Our relationship: [E.G. SUPPLIER WE WANT TO KEEP / COLLEAGUE IN ANOTHER TEAM].
Write an email of no more than 120 words that assumes good faith, restates the request in one sentence, gives the deadline, and makes it easy to reply with a yes or a date. No apologies for chasing, no exclamation marks, no "just checking in". Tone: [WARM AND DIRECT].

Read it as the recipient

Before you edit, paste the draft back and ask the tool to read it as the person receiving it: what lands first, what grates, what is missing, and what question they will fire back. This catches the sentence that sounds fine to you and defensive to them. It takes thirty seconds, and it is the step most people skip.

Read this email as if you were [RECIPIENT'S ROLE AND SITUATION, E.G. A SUPPLIER WHO IS ALREADY BEHIND AND STRESSED]: [PASTE THE DRAFT].
Tell me, in four short bullets: what you understand you are being asked to do; the sentence that would irritate or worry you most, and why; anything you would need to know before you could act; and the first question you would reply with. Do not rewrite the email.

3.2 From meeting notes to action lists, decisions and status updates

You walk out of the meeting with a page of half-sentences, or a transcript nobody will ever read in full. By Thursday two of the actions have evaporated, one decision is being re-argued, and the "quick update" for your director takes forty minutes to write. This is the most mechanical writing you do, and the easiest to hand over, provided you are precise about what you want back.

Three piles, not one summary

Do not ask for "a summary". Ask for three separate lists, because they do different jobs:

• Decisions — what was actually agreed, and by whom. Tell the tool to separate things that were decided from things that were merely suggested; left alone, it will promote a passing idea into a commitment.
• Actions — each with an owner and a date. Insist that it writes "none given" where no owner or date was mentioned, rather than guessing one. The gaps are the most useful part of the output.
• Open questions — anything raised and not resolved, with who needs to answer it.

Ask it to mark anything it inferred rather than read, so you can check those lines first.

Callum, a project lead at Brightwater Property in Manchester, ran a fifty-minute site-meeting transcript through the prompt below. He got nine actions. Two had no owner, and one date was the tool's guess, flagged as such. Instead of inventing owners, he put the two unowned actions at the top of his follow-up email as questions. Both were claimed before lunch, which is more than the previous month's minutes had managed.

Here are my notes (or the transcript) from [MEETING NAME, DATE, WHO ATTENDED BY ROLE]: [PASTE NOTES].
Produce three lists.
1. Decisions: only things that were clearly agreed. Write "suggested, not decided" for anything proposed but not confirmed.
2. Actions: a table with the columns Action, Owner, Due date, Source line. Where no owner or date was given, write "none given" rather than guessing.
3. Open questions: unresolved points and who needs to answer each one.
Mark anything you inferred rather than read with [INFERRED]. Do not add actions that were not discussed. Use names and terms exactly as they appear in the notes.

The weekly status update

A good status update has four headings and fits on a phone screen: Done, Next, Blocked, Decisions needed. Paste last week's update and this week's notes, and ask for the new one in the same format and length. Two things to watch. The tool will soften "Blocked" into "in progress" unless you tell it not to, and "Decisions needed" must name the decision, who makes it and by when, or it is just a complaint.

Once you have the format, it becomes a template: same prompt, new notes, every Friday. Your director learns where to look, and you stop rebuilding the structure each week.

Write my weekly status update for [PROJECT OR TEAM] for the week ending [DATE], for [READER, E.G. MY DIRECTOR]. Use exactly four headings: Done, Next, Blocked, Decisions needed.
Here is last week's update, for format and length: [PASTE LAST WEEK'S UPDATE]. Here are this week's notes: [PASTE NOTES].
Rules: keep it under 150 words. Under Blocked, state plainly what is stuck and since when, without softening. Under Decisions needed, name the decision, who makes it and the date it is needed by. Do not include anything that is not in the notes.

3.3 Reports and proposals: outline first, section by section, executive summary last

You have a proposal due Friday, so you type "write a proposal for..." and get fifteen hundred words of smooth, confident filler. Every paragraph is plausible and none of it is yours, so you spend the afternoon dismantling it. The problem is not the tool. It is that you asked for the whole building before you had agreed the plans. Long documents want a different rhythm: outline, then sections, then the summary, then a tidy-up.

Step one: the outline, fixed before any prose

Brief the tool with CRAFT as usual, but ask only for an outline: each section with a one-line purpose (what it must prove or convey to this reader) and a note of which source material it will draw on. Then do your real thinking here. Delete the section that is there out of habit, merge the two that say the same thing, move the recommendation up. Ten minutes on the outline saves an hour on the prose, because every later change to the structure means rewriting whole sections.

I am writing a [REPORT / PROPOSAL / BUSINESS CASE] for [READER AND THEIR ROLE] about [SUBJECT]. The reader's decision or question is: [WHAT THEY NEED TO DECIDE]. Length target: about [NUMBER] words. Source material I will provide: [LIST, E.G. DISCOVERY CALL NOTES, LAST YEAR'S FIGURES, THE CLIENT'S BRIEF].
Give me an outline only, no prose. For each section: a working title, one sentence on what it must establish for this reader, which source it draws on, and a rough word count. Then list the two sections you think are weakest or missing, and ask me any questions you need answered before drafting.

Step two: one section at a time, with the source pasted in

Draft each section in its own request, with the relevant material pasted in: the client's own words for the context section, the figures for the costs section, your notes for the approach. Ask for that section only, at the agreed length, and tell the tool to write [NEEDS: what is missing] wherever it would otherwise have to invent a fact. A gap you can see is far cheaper than a confident sentence you have to find later.

Mei, a consultant at Kōwhai Consulting in Wellington, used this on a proposal for a regional council. Her outline came back with seven sections; she cut it to five. For "Our approach" she pasted the discovery-call notes and got a usable draft. For "Pricing" she wrote the section herself and used the tool only to check that it read consistently with the rest.

Step three: the executive summary, last

The summary is written last because only then do you know what the document actually says. Paste the finished sections and ask for a summary a busy reader could act on without reading further: the recommendation, the two or three reasons, what you are asking for, and the cost or time involved. If the summary is hard to write, the document is not finished.

Here is the full draft of my [DOCUMENT TYPE] for [READER]: [PASTE ALL SECTIONS].
Write an executive summary of no more than [WORD LIMIT] words that a reader who reads nothing else could act on. Order: the recommendation in the first sentence; the two or three reasons that matter most to this reader; exactly what we are asking them to approve or decide; the cost, timing and main risk. Use only what is in the draft. If the draft does not clearly support a statement, leave it out and tell me separately what is missing.

Step four: the consistency pass

Finally, paste the whole document and ask for a list of inconsistencies, not a rewrite: terms used two ways (client in one section, customer in the next), numbers that disagree between sections, a tone that shifts from formal to chatty, headings that no longer match the outline. You fix each one yourself, so the document stays in your voice and you know exactly what changed.

3.4 AI as a thinking partner: pre-mortems, devil's advocate, decision memos

The expensive part of a bad decision is usually the argument you never had. The colleague who would have spotted the flaw was in another meeting; the finance director asked the awkward question after you had committed; nobody wanted to be the one who said the plan would not work. A chat tool will argue with you at four in the afternoon, tirelessly and without office politics. It argues from patterns, not from knowledge of your situation, so treat everything it says as a question to answer, not a verdict to accept.

The pre-mortem

Describe the plan, then say: "It is six months later and this has failed. Tell me why." Ask for the story first, in the voice of the person writing the review afterwards, and then for the three most likely causes with one early warning sign for each. The time-shift matters. "What could go wrong?" produces a polite list; "why did this fail?" produces specifics, because the tool is now explaining rather than speculating.

Fatima, who runs operations at Maple Ridge Accounting in Calgary, used it before moving clients onto a new document portal. The failure story centred on the fortnight around the tax deadline, when staff would have no time to help clients who could not log in. She had planned the launch for exactly that fortnight. She moved it.

Here is a plan I am about to commit to: [DESCRIBE THE PLAN IN FIVE TO TEN LINES: WHAT, WHO, WHEN, BUDGET OR EFFORT, WHAT SUCCESS LOOKS LIKE].
It is now six months later and the plan has failed. Write a short account, in the voice of the person conducting the review, of what went wrong and why, with specific and realistic causes rather than generic risks. Then list the three most likely causes, and for each one give an early warning sign I could watch for in the first month. Base your reasoning on what I have told you; where you are assuming something about my organisation, say so.

Argue against my plan

Next, ask it to make the strongest possible case against the plan, with no attempt at balance, and then to say which of its own arguments it finds most convincing. Your job is to rebut each one in writing. The argument you cannot rebut is your homework for the week, and far better found now than in the meeting.

The decision memo

When you need to recommend something, ask for a memo in a fixed shape: the situation in three lines, three genuine options (including "do nothing"), the criteria that matter to your reader, a table scoring options against criteria, your recommendation, and what would change your mind. The tool drafts the skeleton; you fill in the cells you actually know and ask it to mark the ones it guessed. Fatima's table looked like this:

Then ask the question that tests the whole thing: "What would a sceptical finance director ask about this before approving it? Give me the five hardest questions and do not answer them." Answer them yourself, in writing, before anyone else gets the chance to ask.

3.5 ChatGPT vs Claude for writing, and Notion AI inside your notes

Someone on your team will ask which one is "best". For writing, the honest answer is that both are good general writers and either will do everything in this course. The differences are in feel and in what surrounds the writing, and the only comparison that counts is one done on your own documents, by you.

What people tend to notice

Treat these as tendencies to test, not facts to repeat. Many people find Claude particularly strong on long, nuanced documents: following a detailed brief and a voice file closely, holding a lot of pasted material in mind, and keeping a careful tone through a difficult message. Many find ChatGPT the stronger everyday generalist, with a broad set of extras around the writing, from working with images and voice to a large library of ready-made custom assistants, and a quick conversational back-and-forth. Both change often, so anything more specific than this would be out of date before you finished the course, which is why you will find no version numbers or league tables here.

The only test that matters

Take three pieces of writing you actually produce: say a client email, a status update and a one-page summary. Give both tools the identical brief, with the same voice file, and then notice one thing: which draft you edit less. Not which sounds cleverer, not which is longer. Editing is the honest measure, because every sentence you change is time the tool did not save you.

Liam, a marketing manager at Northwind Logistics in Leeds, ran this for a week. For short customer emails he barely edited either. For the quarterly review he edited one noticeably less than the other, and that became his tool for long pieces. His team's answer to "which one is best" is now "for what?", which is the right answer.

I am testing how well you write in my style. Here is my voice file: [PASTE VOICE FILE]. Here is the brief. Context: [TWO SENTENCES ON THE SITUATION]. Task: [WHAT THE PIECE MUST DO FOR THE READER]. Format: [LENGTH AND SHAPE]. Tone: [TONE]. Source material to use: [PASTE NOTES OR FACTS].
Write the piece. Then, in a separate short section, list any assumptions you made about facts that were not in the material. Do not use filler openings or closing pleasantries.

Notion AI: working where your notes already live

If your team keeps its notes, project pages and wiki in Notion, the assistant inside it has one advantage over a separate chat window: it works on the page in front of you, without copying anything out. The uses that pay off quickly are plain ones.

• Summarising and tidying — turning a long, messy meeting page into headings and short paragraphs, or putting the key points at the top so people actually read it.
• Page to actions — asking for the actions, owners and dates from a page, as in lesson 3.2, and turning them into tasks that live alongside the project.
• Keeping a team wiki consistent — asking it to rewrite a new page in the style of your existing ones, or to point out where two pages describe the same process differently.

The same cautions apply as everywhere else today. It knows only what is in your workspace plus its general knowledge, so verify anything factual it adds. And check whether your employer has approved the AI features of the tools you already use; a familiar workspace is not automatically an approved one.

4.1 Perplexity for sourced research you can actually cite

It is late afternoon and your director wants "a quick look at what other firms are doing about this" before tomorrow's meeting. You could spend an hour in a search engine opening tabs. Or you could ask a chat tool and get a confident, tidy answer with no way of knowing where any of it came from. You already know from the lesson on hallucinations why the second option is risky: a chat tool answering from memory will invent a report title or a figure as readily as it recalls a real one.

An answer engine is built for exactly this gap. Perplexity is the one we teach, but the habits below work with any tool that searches first and shows its sources.

What an answer engine actually does

When you ask a question, it runs a web search, reads the pages it finds, and writes a short answer with numbered references pointing back to those pages. That is the difference from a plain chat tool: the answer is anchored to things you can open. The summary is still written by a language model, so it can misread a page, lean on a weak one, or blend two sources into a sentence neither of them said. Treat the sources as the product and the summary as a guide to them.

Asking well: four parts

• The question — one sentence, specific enough that a colleague could not misunderstand it.
• The scope — which country or region, which industry, what size of organisation. "In Australia, for small property managers" changes the answer completely.
• The time frame — "published in the last two years", or "since the rule change in [year]". Without it you get current and stale material presented as one picture.
• The kind of source you want — government or regulator pages, industry bodies, named publications, peer-reviewed research. Say what you do not want too, usually vendor marketing pages.

Open the sources. Every time.

Priya, marketing lead at Brightwater Property in Brisbane, asked for a brief on how property managers are using virtual tours. She got a readable answer with seven sources. When she opened them, two were blog posts by companies selling virtual-tour software, one was a newspaper piece several years old that the answer had described as recent, and one did not say what the summary claimed at all. Three sources survived. Her brief cited those three, and said so.

For each source, check three things: does the page exist and say what the answer says it says; who published it and what they want from you; and when it was written. A source you have not opened is not a source. It is a rumour with a link attached.

Keep a sources note as you go

Open a plain document and, for every source you keep, record the link, the publisher, the date, the one sentence you are relying on, and the date you checked it. Most answer engines let you save threads or group them by topic, which is handy, but your reader cannot see your threads and you may not find them next month. The note travels with the brief.

Where it is weak

• Paywalled material — articles behind a subscription. It reads whatever preview is free and may present that fragment as the whole piece.
• Very recent events — search takes time to catch up with new pages, and the first reports of anything are often wrong.
• Niche local topics — a small council's decision, a regional supplier's terms. With little on the web, it reaches for whatever exists, which is often about somewhere else.
• Numbers that disagree across sources — it may quietly pick one. Ask it to show each source's figure separately so you can see the spread.

I need a short research brief for work.
Question: [THE QUESTION IN ONE SENTENCE].
Scope: [COUNTRY OR REGION, INDUSTRY, SIZE OF ORGANISATION].
Time frame: only use material published since [MONTH AND YEAR]. If nothing that recent exists, say so rather than using older material.
Preferred sources: [GOVERNMENT OR REGULATOR SITES / INDUSTRY BODIES / NAMED PUBLICATIONS]. Avoid vendor marketing pages unless you label them as such.
Format: five to eight bullet points, each ending with its source number, then a list of the sources with publisher and publication date.
Where sources disagree, show each position or figure separately rather than averaging or choosing one.
Finish with a list of anything you could not find or could not confirm.

4.2 Claude and Gemini for long contracts, reports and spreadsheets

A forty-eight-page supplier contract arrives at three o'clock. The supplier call is tomorrow and your director wants a recommendation by Friday. You are not going to read every word. What usually happens next is skimming, hoping, and deciding on the four pages you happened to read.

The problem is not reading speed. It is not knowing which pages matter. This lesson is about getting a long document to tell you that.

Which tool, and the first rule

Claude and Gemini both take long documents, uploaded as a file or pasted as text, and hold the whole thing in view while they answer. ChatGPT handles uploads too; the method is the same. The first rule comes from the confidentiality lesson: a contract, a tender or a board report is confidential by default, so use only the tool your employer has approved for that material. If there is no approved tool, this lesson waits.

Ask for decisions, not a summary

"Summarise this" gives you the document's own structure in fewer words, which is pleasant and useless. "Summarise this into the decisions I need to make" gives you an agenda. Tell the tool who you are, the deadline, and what is at stake in money, people or risk, so the decisions it finds are the ones that matter at your level.

Pull the clauses into a table

The second pass is a table with four columns: the clause, what it says in plain English, the risk to you, and the question to ask. The last column is the useful one. It turns a wall of legal text into a one-page list of questions for the other side or your adviser, and stops you paying a lawyer to read what you could have read yourself.

Compare two versions

When a revised draft arrives, upload both versions and ask for every change, old wording and new, with a flag on anything that moves risk or money. Your word processor's own compare feature is more reliable at finding every change, especially one altered number. AI is better at telling you what the changes mean. Use both, in that order.

Questions that must come with a quote

Any follow-up question gets the same condition: answer only from the document, quote the exact passage, give its clause number, and if the answer is not there, say so. That condition turns the tool from a storyteller into an index. It cannot invent a clause it has to show you.

Then verify against the source

Open the original. For each quoted passage, search for a distinctive phrase and check that the quote is real and the plain-English version faithful. Look hardest at three things: numbers (notice periods, amounts, dates), who the obligation falls on (you or them), and anything that follows "unless", "except" or "subject to". That is where meaning hides.

I am [YOUR ROLE] at [YOUR ORGANISATION]. I have attached [WHAT THE DOCUMENT IS, E.G. A 48-PAGE SUPPLIER CONTRACT FOR A BOOKING PLATFORM]. I must [RESPOND / SIGN / RECOMMEND] by [DEADLINE]. It matters because [ONE SENTENCE ON THE MONEY, PEOPLE OR RISK INVOLVED].
Do not summarise the document section by section. List the decisions I personally need to make or escalate, most important first. For each one give: what I am deciding, the clause or page it comes from, what the document currently says in one plain-English sentence, and what I would need to know before deciding.
Use only what is in the document. If you are unsure about anything, say so rather than guessing. Keep the whole answer under 300 words.

Using only the attached document, build a table with four columns: Clause (number and heading); What it says (one or two plain-English sentences); Risk to us (low, medium or high, with a one-line reason); Question to ask (for the other party or our adviser).
Cover every clause about: price and price changes, term, renewal and termination, liability and indemnity, data and confidentiality, service levels, and anything that lets the other party change the terms on their own.
Quote the exact wording for anything you mark high risk. If one of those topics is not covered anywhere in the document, add a row saying so, because silence can be a risk too.

4.3 Spreadsheets and data without the panic

An export lands in your inbox with a question attached: "Can you see why travel spend jumped in the spring?" You know enough to filter and add things up, and not much more. The honest options used to be an afternoon of guessing or a favour from someone in finance.

AI can do a lot of this with you, but data has sharper edges than prose. A wrong formula produces a number that looks exactly like a right one. So this lesson is as much about checking as asking.

Two routes

The first route is the AI built into your spreadsheet tool: Gemini in Google Sheets, Copilot in Excel. They can see the sheet you are working on, write a formula from a plain-English request, suggest a chart and explain a formula someone else wrote. The data stays where it is, which is the main advantage.

The second route is pasting a sample into ChatGPT or Claude: the header row plus twenty or thirty rows, or an uploaded CSV file (a plain-text version of a spreadsheet that any tool can read). This is better for thinking out loud, learning what the formulas do and planning the clean-up. It also means the data has left your spreadsheet, so the privacy rule below applies in full.

Describe the data before you ask anything

Most bad answers come from the tool guessing what a column means. Tell it what one row represents ("one row is one expense claim"), roughly how many rows there are, what each column is and means, and what question you want answered. "Date" could be the date of the expense, the claim or the payment. You know which. The tool does not. This is the Context part of your CRAFT brief, and with data it is most of the work.

Ask for three things

• Cleaning steps — blanks, duplicates, the same category spelt three ways, dates stored as text, currency symbols sitting inside number cells. Ask for the steps and the order to do them in.
• Formulas with an explanation — a formula you cannot explain is one you cannot fix or defend. Ask for one line on what each part does.
• Chart suggestions — ask which chart fits your question and why, not just "make a chart". A change over time wants a line; a comparison between a few categories wants bars; many pie charts should have been a table.

Sanity-check before you believe it

Recompute the headline total with a plain SUM and compare it with the source system's own report. Pick three rows at random and trace them through by hand. Count the rows before and after cleaning. Look at the largest value, the smallest and the blanks. If a result surprises you and you cannot explain why, you have found either an insight or an error, and you must find out which before anyone else sees it.

Jordan, a finance assistant at Maple Ridge Accounting in Calgary, did exactly this with a client's fourteen months of expenses, after removing the column of employee names. The clean-up revealed "Travel", "travel " and "TRAVEL-Air" as three separate categories, which alone explained most of the apparent jump. The chart the tool suggested, one line per category by month, showed the rest came from a single conference. Jordan's total matched the client's system to the dollar before the chart went anywhere.

I am working with a spreadsheet export. One row is [WHAT ONE ROW REPRESENTS, E.G. ONE EXPENSE CLAIM]. There are about [NUMBER] rows in total. The columns are: [ONE LINE PER COLUMN: NAME AND WHAT IT MEANS, E.G. "Date = date the claim was submitted"]. No column identifies a person.
Here is a sample including the header row: [PASTE 20 TO 30 ROWS].
I want to know: [THE QUESTION IN ONE SENTENCE].
Please: 1) list the cleaning problems you can see in the sample and the steps to fix each, in order; 2) give me the formulas to answer the question in [GOOGLE SHEETS / EXCEL], with a one-line explanation of what each part does; 3) suggest one chart and say why it fits the question; 4) tell me which checks to do by hand before I trust the result.

4.4 The five-minute verification routine before anything leaves your desk

Everything you have done today produced something that reads well. That is the danger. Your proofreading instinct was trained on human mistakes, which look like mistakes: a typo, a clumsy sentence, a number with too many zeros. AI mistakes look like good writing. A made-up clause number sits in a perfectly formed sentence. You cannot rely on noticing. You need a routine that does not depend on it.

This is that routine. Five checks, five minutes, every time something AI helped with leaves your desk. Learn the five words and you will not need this lesson again: Sources, Sums, Specifics, Sceptic, Say so.

1. Sources: where did each fact come from?

Go through the draft and sort every factual statement into three piles: from material you gave the tool, from a source you opened yourself, or from the model's general knowledge. The third pile is your work. Each item in it gets checked against a real source or cut. "I think that's right" does not move anything out of the pile.

2. Sums: recompute or cross-check every number

Every total, difference, duration and date calculation. Recompute it yourself or check it against the system it came from. Pay special attention to numbers the tool derived rather than copied: "a rise of a fifth", "three weeks later", "the cheaper of the two". Those are where plausible and true part company.

3. Specifics: names, dates, titles, references

Every person's name and job title, every organisation, every clause number, every date. Is it spelt right, is it the right person, does the date exist and fall on the day the draft says it does. This check takes one minute and catches the mistakes that embarrass you most.

4. Sceptic: what would make this wrong?

Spend one minute as your own hostile reader. What assumption is this built on? Who in the room would disagree, and what would they say? What is missing that the reader will expect? You can ask the tool to play this part too; it is good at criticising its own work when asked plainly. Then decide whether the draft survives.

5. Say so: disclose when it is appropriate

Your employer's policy decides when you must say that AI was involved; follow it. Where it is silent, ask whether the reader would want to know, which is usually yes for anything external, formal or regulated, and for anything they will rely on without checking. A plain sentence does the job: "Prepared with AI assistance and checked against the source documents by [your name]." Disclosure is not an apology. Checked work with a disclosure is worth more than unchecked work without one.

Here is a draft that is about to go to [WHO WILL RECEIVE IT AND WHY IT MATTERS]: [PASTE THE DRAFT].
Do not rewrite or improve it. Produce a verification checklist I can work through in five minutes, in five sections:
1. Sources: list every factual claim and mark each as "from material I gave you", "from a source I named" or "from your general knowledge".
2. Sums: list every number, date and duration, and show me how to check each one.
3. Specifics: list every name, title, organisation, clause or reference I should confirm.
4. Sceptic: give the three most likely ways this draft could be wrong or misread, and the first question a hostile reader would ask.
5. Say so: tell me whether this is the kind of document where I should state that AI assisted, and suggest one sentence of wording.
One line per item.

5.1 Writing a visual brief that gets what you meant

You have probably typed "a photo of our product on a desk" into an image tool and received something that was nobody's desk, in nobody's office, holding a product that was nearly yours. The tool did not fail. It filled in every detail you left out, and it filled them in with the most average choice available.

Day 2 taught you to brief AI like a colleague. The same idea applies here with one twist: a photographer needs different information from a writer. Give an image tool the seven things a photographer would ask you before picking up a camera, and the results stop being a lottery.

The seven parts of a visual brief

• Subject — what is in the picture, with the specifics that matter: "a 30 ml amber glass dropper bottle with a black cap", not "a serum".
• Setting — where it sits, what it sits on, what is behind it: "a pale timber bathroom shelf, white tiles out of focus behind".
• Style or medium — photograph, flat illustration, watercolour, 3D render. If photograph, which kind: a clean studio shot, a lifestyle scene with the product in use, or editorial, the magazine look.
• Mood and lighting — where the light comes from and how it feels: "soft morning light from a window on the left, warm, slightly hazy".
• Composition and framing — eye level or overhead, close-up or wide, and where you need empty space for text later: "product on the right third, empty space on the left".
• Colours — two or three, in words: "pale sand, deep green, a touch of brass". Some tools also accept the six-character colour codes your designer uses; words are the safer bet.
• What to avoid — "no text, no people, no extra products, no glossy reflections, nothing floating".

Then add one line that never changes: your brand line. "Our look is clean and warm, natural materials, pale sand and deep green, no gradients, no neon." Paste it into every brief and your images will belong to the same family even when the subjects differ.

Why text inside images goes wrong

Image tools draw letters as shapes, not as words. Your brand name comes back with an extra letter, your tagline contains a word that does not exist, and the font is whatever the tool felt like. Some tools manage a short word or two reasonably; none are reliable enough to publish without reading every letter. The dependable method is to generate the image with a clear empty area and add the words afterwards in Canva, where you choose the font and the spelling is yours.

A [PHOTOGRAPH / FLAT ILLUSTRATION / 3D RENDER] of [SUBJECT WITH SPECIFIC DETAILS: SIZE, MATERIAL, COLOUR, SHAPE].
Setting: [WHERE IT SITS, WHAT IT SITS ON, WHAT IS BEHIND IT].
Lighting and mood: [WHERE THE LIGHT COMES FROM, HOW IT FEELS].
Composition: [EYE LEVEL / OVERHEAD / CLOSE-UP], subject on the [LEFT / RIGHT / CENTRE], empty space on the [SIDE] for text added later.
Colours: [TWO OR THREE COLOURS IN WORDS].
Avoid: no text, no logos, no people, [ANYTHING ELSE TO EXCLUDE].
Our look: [ONE SENTENCE ON YOUR BRAND'S VISUAL STYLE].

5.2 Higgsfield for product shots and social images

You have a product, a launch date and no photographer booked. Or you have one decent photo and need a month of posts that do not all look like the same photo. Higgsfield is built for this: it produces still images and short video from a text brief or from a photo you give it, and it works best when you treat it as a very fast studio rather than a magic button.

The workflow, start to finish

1. Start from your brief or a reference photo. A text brief alone works. Better, when you have one, is a plain photo of your real product on a neutral background, uploaded as the starting point, with your brief describing the scene around it. The product then stays your product instead of a close cousin.
2. Generate several, not one. Ask for a handful of variations at once. Compare them side by side and pick the one closest to your brief, not the prettiest. Pretty-but-wrong costs you the next twenty minutes.
3. Refine one element at a time. Keep the brief identical and change only the background, or only the lighting, or only the angle. Change three things and you will not know which one helped.
4. Export in the sizes you need. Decide the formats before you finish: square for feeds, tall for stories, wide for a web banner. Either generate each format or make sure the composition survives a crop, which is why you left empty space in the brief.
5. Save what worked. Keep a small library: the brief, the tool, the date and the final image, in Notion or a plain document. Next month starts from a brief you already know works, not from a blank box.

What to expect, honestly

Hands come out with the wrong number of fingers or impossible joints. Small text is garbled. Logos are approximations of your logo. Exact product details drift: the cap is a slightly different shape, the label is blurred, the bottle is the wrong shade of amber. None of this is a reason not to use the tool. It is a reason to know which images are fine as they are and which need a real photograph or a fix in Canva.

Generated images also tend to look generated: too glossy, too perfect, colours turned up too far. Asking for "natural, slightly uneven light, realistic textures, not glossy" in the brief usually helps.

Use the attached photo as the product. Keep its shape, proportions, colours and label exactly as shown; do not redesign it.
Place it in this scene: [SETTING, SURFACE, BACKGROUND].
Lighting and mood: [DIRECTION, WARMTH, SOFT OR HARD].
Composition: [ANGLE], product on the [LEFT / RIGHT / CENTRE], clear empty space on the [SIDE] for text added later.
Colours: [TWO OR THREE COLOURS IN WORDS].
Style: natural photograph, realistic textures, slightly uneven light, not glossy, not over-saturated.
Avoid: no text, no extra products, no hands, no reflections of people.
Our look: [YOUR BRAND LINE].

5.3 Higgsfield video: short clips and ads from a text or an image

Your director wants "a short video for the launch", there is no budget for a crew, and the deadline is next week. Not long ago that meant stock footage and an apology. Now it means a still photo of your product, a few sentences describing how it should move, and an afternoon of trying things.

Short video does not need to be clever. It needs to make someone stop scrolling for a moment, understand what they are looking at, and know what to do next. That is a structure, and structures can be briefed.

Three beats

• Hook — the opening moment. Something moves, something appears, or the frame is simply good enough to pause on. Steam rising. A lid lifting. Light sweeping across a surface.
• Product or benefit — what it is, or what it does for the person watching, shown rather than explained. The drink in the cup. The jacket in the rain.
• Call to action — what to do next, usually a final held frame where the words and the logo go later: book, visit, try, order.

For a short social clip each beat is one scene, and each scene is one generation. Generated clips are short, a few seconds each, so a finished ad is usually two or three clips joined in an editing tool rather than one long generation.

Format and length

Decide before you generate. Tall, vertical video is for phones: stories, reels and short-form feeds. Wide, horizontal video is for websites, presentations and anything watched on a laptop or television. Square sits between the two and is forgiving on most feeds. Changing the shape afterwards (reframing) crops the edges, so a product placed at the side of a wide frame can vanish from the tall version. Short social clips run from a few seconds to well under a minute; shorter is almost always better than you think.

Turning a still into motion

Upload the product shot from 5.2 and describe the movement in plain words: "slow push in towards the bottle", "steam rises gently from the cup", "the camera drifts left while the light warms". Small, slow movement looks real. Big movement — people walking, hands pouring, fabric flapping — is where the seams show: objects warp, handles swap sides, text flickers.

Describe the camera like a director. "Static camera, only the steam moves." "Slow orbit around the product." "Handheld feel, slight sway." Then iterate the way you did with images: one change per attempt. If the product distorts, reduce the movement or move the camera instead of the product.

Captions, music and your logo come afterwards, in Canva or whichever editing tool you have. Do not ask the model to render words or logos in the clip; they will not survive. Use music you are licensed to use.

I am making a short [VERTICAL / HORIZONTAL] video ad for [PRODUCT OR SERVICE: TWO SENTENCES ON WHAT IT IS AND WHO IT IS FOR]. The audience is [WHO]. The feeling should be [THREE ADJECTIVES].
Write a three-scene storyboard: hook, product or benefit, call to action. For each scene give me: a one-line description of what is in frame, the single movement that happens, the camera behaviour (static, slow push in, drift, orbit), the lighting, and what should stay completely still. Keep each scene to one movement. No text, logos or people in any scene; I will add words and the logo afterwards. Write each scene so I can paste it straight into an image-to-video tool.

5.4 Canva Magic Studio: keeping everything on-brand

You now have a dozen good images and a clip. The problem shifts: they have to look like they came from the same company as last month's work, in your fonts and your colours, and the AI tool does not know what those are. Generating is the easy half. Consistency is what makes a small business look like a serious one, and it is what Canva is for.

Set the brand once

Canva has a brand kit: a place to store your colours, fonts and logo so every new design starts from them instead of from the defaults. If you have brand guidelines from a designer, put those values in once. If you do not, choose two colours and two fonts, write them down, and treat that as your brand kit. Deciding is most of the work.

Then pick a small set of templates and stay with them: one social post layout, one story layout, one one-pager, one presentation. Repetition is not boring. Repetition is the brand.

What Magic Studio does for you

Magic Studio is the collective name for Canva's AI features. The ones that matter for this week's work:

• Resizing one design into many formats — make the post once, turn it into the story and the banner, then check each one, because text wraps and images crop differently at each size.
• First drafts of presentations and one-pagers — describe what you need or paste your text and the tool produces a structured first version. Treat it exactly like a first draft from a writing tool: the structure is a starting point, the placeholder text needs replacing, and your brand needs applying.
• Image help — generating an image inside Canva, removing a background, extending a photo to fit a wider frame, or editing out an object. Useful for small fixes; for product scenes, start in Higgsfield and finish here.
• Combining AI images with your layouts — drop the Higgsfield image into your template, then put the real logo and the real words in the empty space you left in the brief. This is where "no text in the image" pays off.

The "does it look like us" check

Put the new design next to three older ones you are proud of. Ask three questions. Same colours? Same type? Same photographic feel, warm or cool, busy or clean? Then the hard one: would a customer know it is yours with the logo covered? If any answer is no, fix that one thing before touching anything else.

Callum, who runs operations at Brightwater Property in Bristol, set up a brand kit and two templates so that every agent could make their own listing posts. The check caught the first problem within a week: a post in a bright red that appears nowhere in the brand. The fix was a sentence in the template notes, not a redesign.

A [ONE-PAGE FLYER / SHORT PRESENTATION] for [AUDIENCE] from [COMPANY], titled "[TITLE]".
Sections, in this order: [SECTION 1], [SECTION 2], [SECTION 3], [CALL TO ACTION].
Use our brand colours and fonts. Style: clean, generous white space, one image area per section, no decorative icons, no stock-photo people.
Leave a clear space at the top for our own photograph and a space at the bottom for the logo.
Placeholder text in plain English, no slogans, British spelling.

5.5 Copyright, likeness and disclosure: the rules for commercial use

The image is finished and it is good. Then someone — your manager, a client, the one colleague who reads terms and conditions — asks, "Can we actually use this?" You want to be able to answer before it is published, not after a complaint.

This lesson gives you general principles for the UK, Canada, Australia and New Zealand. It is not legal advice, and your employer's own policy on AI-made content comes first. For anything large, such as a national campaign, packaging or a client's brand, get a professional opinion.

Five principles

1. Read and keep the licence terms of each tool. Whether you may use outputs commercially, who owns them, and whether your uploads may be used to train the tool differ between tools and change over time. Read the terms before commercial use, save a dated copy, and note in your brief library which version you relied on. Never upload an unreleased product or confidential material without knowing what happens to it.
2. Do not generate real people or other people's marks. No celebrities or public figures. No colleagues, employees or customers without written consent, and consent to use a photo is not automatically consent to animate or alter it. No other companies' logos, characters, mascots or packaging, and no "in the style of" a named living artist or brand. Treat faces and voices as belonging to their owners.
3. Label where it is expected. Some platforms ask you to declare AI-made content or add a label themselves; some industries and many clients have their own rules. When unsure, label. A short "image generated with AI" costs nothing and protects trust. Never present a generated image as a real photograph of a real thing when someone would rely on it: a room in a property listing, a dish in a restaurant, a clinical result.
4. Keep a record. The brief, the tool, the date, which reference images you used and where they came from (your own photos or licensed stock, not a search-engine grab), and the final file. If a question arrives in six months, you can answer it in five minutes.
5. Advertising standards still apply. In all four countries an advertisement must be truthful and not misleading, and that applies however the image was made. If the generated picture shows a product doing something the real one cannot, or looking like something it is not, it is a misleading ad, and "the AI did it" is not a defence.

I am about to publish an AI-generated [IMAGE / VIDEO]. Here is the brief I used: [PASTE BRIEF]. Reference images I used and where they came from: [DESCRIBE]. It will appear on [PLATFORM OR PLACE] in [COUNTRY], as [AN ADVERTISEMENT / A SOCIAL POST / A CLIENT DELIVERABLE].
Check it against this list and mark each item clear, check or stop, with one line of reasoning: real people or likenesses; third-party logos, characters, brands or named artists' styles; anything shown that the real product or service cannot do; whether an AI label is expected in this context; whether the reference images were ours or licensed. Finish by reminding me that this is not legal advice and that my employer's policy takes priority.

6.1 Reusable assistants: custom GPTs and Claude Projects

Every Friday you open a fresh chat and type roughly the same thing: who you are, what the report is for, how long it should be, the headings your director expects. Then you paste your voice file. Then you paste last week's version so the tool knows the shape. Fifteen minutes of set-up for a task that is supposed to save you time.

The tools now let you do that set-up once. ChatGPT calls the result a custom GPT, Claude calls it a Project, and Gemini has an equivalent. The idea is the same in each: a space with standing instructions and reference files that are already loaded every time you open it.

When a task deserves its own assistant

Not every task does. Ask three questions:

• Do you do it at least weekly, or do several people on your team do it?
• Is the brief always the same, with only the raw material changing?
• Is the output format fixed: the same headings, the same length, the same audience?

Three yeses means you are re-typing the same brief every time, and that brief belongs in an assistant. One no means a saved prompt in your library is enough.

The three building blocks

1. Standing instructions — your CRAFT brief, written once: role, context, task, format, tone, plus the rules you keep having to repeat ("never invent figures", "flag anything you are unsure about").
2. Reference files — your voice file, the template the output must follow, and any policy or glossary the tool needs. They live in the assistant, so you never paste them again.
3. Two or three example outputs — real, finished pieces you were happy with, with anything sensitive removed. Examples teach format and tone faster than any description.

A worked example: the weekly status report assistant

Aroha is an operations manager at Kōwhai Consulting in Wellington. Every Friday she turns a week of messages, meeting notes and a project tracker into a one-page status report for the partners. The format never changes: progress, risks, decisions needed, next week.

She built one Project. Its standing instructions are the prompt card below. She uploaded her voice file, the partners' report template, and two reports from earlier in the year that the partners had praised. Now each Friday she opens the Project, pastes the week's raw notes, and gets a draft in the right shape.

She still reads every line, because the tool will still misread a date or soften a risk she wants stated plainly. What has gone is the set-up, and the drift in quality that came from briefing it slightly differently each week.

Keep it alive

Assistants rot quietly. The template changes, a policy is updated, a project ends, and the files inside still describe the old world. Put a monthly reminder in your calendar: open each assistant, remove stale files, replace them with current ones, and re-read the instructions. If you have not used an assistant for two months, delete it. A stale assistant gives confident, outdated answers.

You are my assistant for writing the weekly status report I send to [AUDIENCE, E.G. THE PARTNERS] at [ORGANISATION]. I am [YOUR ROLE].
Each week I will paste raw notes, messages and tracker extracts. Turn them into a report that follows the attached template, in the voice described in the attached voice file. Match the attached examples for length and tone.
Rules: use only facts from the material I paste. If something is missing or unclear, list it under "Questions for me" rather than guessing. Never invent dates, figures or names. State risks plainly, without softening. Keep the whole report under [WORD LIMIT] words.
Finish with a three-line summary I can paste into a chat message.

6.2 Zapier basics: inbox to summary to task list, hands-free

Some of your work is not thinking. It is moving. An email arrives, you read it, you write down what it asks of you, you put that in your task list. Several times a day. None of those steps needs you, but every one of them interrupts you.

Zapier is a tool that connects the apps you already use and passes information between them without you touching it. You do not write code; you fill in a form. This lesson gives you the vocabulary and one safe example.

Triggers and actions, in plain English

Every automation (Zapier calls each one a Zap) is a sentence: "When this happens, do that."

• The trigger is the "when": a new email arrives with a certain label, a form is submitted, a row is added to a spreadsheet, a file lands in a folder.
• An action is the "do that": create a task, send a message, add a row, save a draft email.
• An AI step sits between them. It can summarise, extract, classify or draft, using the same kind of prompt you have written all week. Zapier has its own built-in AI actions and can also connect to ChatGPT or Claude.

Each step takes the output of the previous one as its input, like passing a form along a row of desks.

A three-step example

Mei runs client onboarding at Maple Ridge Accounting in Calgary. Clients email documents and questions in long, rambling messages, and she used to copy the action points into her task tool by hand.

1. Trigger: a new email arrives in her work inbox from a sender in her client list, or carrying the label she applies to client mail.
2. AI step: the email body is passed to an AI action with the prompt card below. It returns a two-sentence summary and a list of actions, each with a deadline if one was mentioned.
3. Action: a task is created in whichever task tool she already uses, titled with the client's name, with the summary and actions in the description.

Nothing goes back to the client. Mei still reads the task before she acts on it. What has disappeared is the copying.

Guardrails

• A human review step for anything outward. An automation that writes to clients should save a draft for you to approve, never send directly.
• No sensitive data through consumer tools. Payroll, health, account details, anything the confidentiality rules cover: keep it out, or use only tools your organisation has approved.
• Test with dummy data first. Send yourself a made-up client email and watch what happens before you connect the real inbox.
• Name and document every automation. A clear name such as "Client email to task", plus one paragraph on what it does and who owns it. In six months you will not remember.

When not to automate

Rare tasks: if it happens once a month, the set-up costs more than it saves. Judgement-heavy tasks: anything where the right answer depends on reading the room, the relationship or the politics. Automate the carrying, not the deciding.

You are an assistant that turns incoming client emails into task notes for [YOUR ROLE] at [ORGANISATION].
Here is the email: [EMAIL BODY FROM THE PREVIOUS STEP].
Write:
1. A summary of no more than two sentences.
2. A list headed "Actions for me", one per line, each starting with a verb. Include a deadline only if the email states one; otherwise write "no deadline given".
3. A line headed "Needs a human decision" if the email contains a complaint, a request to change fees or scope, or anything legal. Otherwise write "none".
Use only what is in the email. Do not add advice, assumptions or pleasantries.

6.3 Microsoft 365 Copilot and Gemini in Workspace, if your company already pays

You may not need to go anywhere. If your organisation runs on Microsoft 365 or Google Workspace, there is a fair chance an AI assistant already sits inside your email, your calendar, your documents and your spreadsheets, and that it is the only one your IT team has formally approved.

This lesson is short on wonder and long on practicality: what these built-in assistants are genuinely useful for, why organisations tend to prefer them, and how to find out what you are actually allowed to use.

What they are

Microsoft 365 Copilot is the AI assistant inside Outlook, Teams, Word, Excel and PowerPoint. Gemini in Workspace is the equivalent inside Gmail, Google Meet, Docs and Sheets. Both are switched on by whoever administers your organisation's account, which is why the lesson title says "if your company already pays": in most workplaces you cannot turn them on yourself.

Their advantage is not that they are cleverer than ChatGPT or Claude. It is that they can see the thread, the meeting or the document you are already in, without you pasting anything.

Practical uses that earn their place

• Email: recap a long thread you were copied into late; draft a reply in your tone that you then edit.
• Meetings: a summary of a recorded or transcribed meeting, with decisions and actions, for the people who were not there.
• Documents: a first draft from your rough notes in the template you are already using; a tighter version of the paragraph you have been staring at.
• Spreadsheets: explain what an inherited formula does in plain English; suggest a formula for a result you describe; point out a column that looks inconsistent.

Fatima is an HR business partner at Harbourline Dental, a group of practices in Melbourne. Her firm switched on Copilot, so she now asks it to recap the fortnightly practice-managers' call and draft the follow-up email. She edits the tone and checks every action against her own notes before sending. Liam, a lettings manager at Brightwater Property in Bristol, uses Gemini in Sheets to explain the formulas in a rent-arrears tracker that nobody admits to building.

Why organisations often prefer them

This varies, so take it as a general picture. These assistants work inside the company's existing environment: the same accounts, permissions and retention policies that already govern your email and files. The vendors position them as keeping your organisation's data within that boundary rather than treating it like a public chat. That is why an IT or compliance team that bans consumer AI tools may still allow these. Whether that holds for your organisation depends on how it has been set up, and only your IT team can tell you.

You still verify everything

A built-in assistant hallucinates like any other. It will summarise a meeting and attribute a decision to the wrong person, or recap a thread and miss the one message that reversed everything. Treat its output as a first draft from a quick but careless colleague, and run the verification routine from Day 4 before anything leaves your desk.

How to find out what is approved

Ask three questions, in order: does our organisation have an approved AI tool, which tasks and which kinds of data is it approved for, and where is that written down. Try your IT help desk, your manager, and any acceptable-use or data policy you were given when you joined. If nobody can answer, that is useful information in itself: be conservative, and apply the Day 1 confidentiality rules in full.

Summarise this email thread for me in five bullet points: what was asked, what has been agreed, what is still open, who owes what, and any deadline mentioned. Quote the exact wording of any commitment made.
Then draft a reply from me to [RECIPIENT] that [WHAT THE REPLY SHOULD DO, E.G. CONFIRMS THE DATE AND ASKS FOR THE SIGNED FORM]. Keep it under [NUMBER] words, in a [CALM AND DIRECT] tone, with no filler phrases.
Mark anything in the reply that you inferred rather than read in the thread with [CHECK].

6.4 Build your first workflow today

You have spent five days using AI inside conversations. The difference between someone who "uses AI" and someone whose week is actually lighter is one small workflow that runs every time without being re-invented. Today you build one. Not the ambitious one; the small, dull one that works.

Step 1: choose the task

Pick something that is all four of these:

• Recurring — weekly or more often.
• Rule-based — you could explain to a new starter how to do it in ten minutes.
• Low-risk — if the output is wrong and nobody catches it, the result is mild embarrassment, not a lost client or a legal problem.
• Annoying — you dread it, which means you will actually maintain the automation.

Priya leads the customer service team at Northwind Logistics in Manchester. Every Monday she downloads a week of customer complaints, reads each one, tags it by theme (late delivery, damaged goods, billing, driver conduct) and counts the themes for the team meeting. Recurring, rule-based, low-risk, annoying. Chosen.

Step 2: map the steps on paper

Write every step as a numbered list, as if explaining it to a temp. Priya's list: download the export, open it, read each complaint, assign a theme, count the themes, write three lines for the meeting, paste them into the agenda. Most people find more steps than they expected. That is the point.

Step 3: decide which single step the AI does

One step. Not the whole chain. Look for the step that is reading or writing, repetitive, and easy to check. For Priya it is assigning a theme, with "write three lines for the meeting" as a possible second phase once the first has proved reliable. Downloading and pasting stay human for now; counting is a spreadsheet's job.

Step 4: build it as an assistant or a Zap

If the raw material arrives in your hands and you are happy to paste it, build an assistant as in 6.1: standing instructions, your theme list as a reference file, two tagged examples. If the raw material arrives in an app and you want the step to run without you, build a Zap as in 6.2, with the same prompt in the AI step. Start with the assistant. It is faster to build and easier to fix.

Step 5: test with three real past cases

Take three weeks you have already done by hand. Run each through the new workflow and compare the result with your own. Where it disagrees, decide who was right. If the tool was wrong, tighten the instructions or add an example, then run it again. Three agreements in a row is your bar for trusting it with next Monday.

Step 6: write the one-paragraph note

What it does, what goes in, what you check before using the output, and who owns it. Save it alongside your voice file. This note is what makes it a workflow rather than a trick you half-remember.

I am [YOUR ROLE] at [TYPE OF ORGANISATION]. Here is a recurring task I do [HOW OFTEN], written as steps: [PASTE YOUR NUMBERED STEPS].
For each step, tell me in one line whether it is best done by a person, by AI, or by a spreadsheet or app rule, and why.
Then recommend the single step where AI would help most, given that it must be low-risk and easy to check. Describe how I would check its output in under two minutes.
List three things that could go wrong with that step and one guardrail for each.
Finally, write the standing instructions I would give an assistant to do that step. Ask me up to three questions first if anything is unclear.

7.1 Assemble your system: prompt library, weekly routine, and the rules for when not to use AI

Look at where this week's work actually lives. Some prompts are in chat histories you will never scroll back through, some are in your worksheets, and one is in a message you sent yourself on Tuesday. That is normal after a week of learning, and it is also how good habits quietly disappear: in a fortnight you are retyping briefs from memory and getting worse results than you did on Day 3.

The fix is small and slightly boring: a library, a routine and a short list of rules.

The library: organise by task, not by tool

Your library is one document, note or spreadsheet that you already open most days, in whatever format your workplace already trusts. Not a new app. Group entries by the job you are doing (writing, research and long documents, analysis, visuals, admin and automation) because that is how you will go looking for them. The tool is a detail you note on each entry.

Each entry has five parts: a name, a one-line "use when", the tool it works best in, the prompt with placeholders in square brackets, and one line on what to check in the output. That last line is the five-minute verification routine from Day 4, attached to the prompt that needs it.

Names start with a verb and say what comes out: "Summarise a long PDF into a one-page decision brief", "Soften a chase email without losing the point". A name like "good email prompt v2" tells you nothing on a busy Thursday.

Aim for ten to fifteen entries you genuinely use, with your voice file from Day 2 pinned at the top. The bonus prompt library holds far more; yours should hold only what has earned its place. Anything you use every day belongs in a reusable assistant, as on Day 6.

The routine: Monday to plan, Friday to prune

Monday, five minutes, before you open anything else: run down the week's tasks and mark which get AI help and which do not. Be specific: "first draft of the board paper, yes; the conversation with Liam about his performance, no". Deciding in advance stops you reaching for the tool out of habit, or avoiding it out of vague unease.

Friday, fifteen minutes: save any prompt that worked, named properly. Move anything you have not touched for three weeks to a "retired" section at the bottom. Write one line about what went wrong this week; that line is what you act on next Monday.

Aroha runs the office at Kōwhai Consulting in Wellington. Her first Friday review turned nine scattered prompts into six named entries, retired two that overlapped, and surfaced one problem: her proposal prompt kept pulling in details from earlier clients' chats. Her note for Monday read "proposal drafts: new chat every time, no previous client material". That is the whole routine working as intended.

The rules: when you do not use AI

Write these as plain statements you could say out loud, and put them at the top of the library, above the prompts. Start here and make the list yours:

• No confidential or personal data in consumer tools — the Day 1 rule. Names, health details, salaries, client financials and anything under a confidentiality agreement stay out unless your organisation has approved the tool for that purpose. UK GDPR, Canada's PIPEDA, Australia's Privacy Act 1988 and New Zealand's Privacy Act 2020 all hold you responsible for personal information wherever you send it.
• No final legal or financial judgement — AI can explain a clause or lay out options; a qualified person decides and signs.
• Nothing you cannot verify — if you have no way to check a claim, it does not go out in your name.
• No relationship-critical messages without a human read — a redundancy conversation, an apology, a complaint reply. Draft if you like; a person reads every word before it goes.
• Nothing your employer's policy excludes — whatever this course says, that document wins.

Here is a messy collection of prompts and notes I used this week: [PASTE EVERYTHING, UNSORTED].
Organise them into a prompt library. For each prompt give me: a short name starting with a verb that says what comes out; a one-line "use when"; the tool I mentioned, or "any" if I did not; the cleaned-up prompt with placeholders in [SQUARE BRACKETS] for the parts that change each time; and one line on what I should check in the output before using it.
Group the entries under these headings: Writing, Research and long documents, Analysis, Visuals, Admin and automation.
Merge near-duplicates and tell me which ones you merged. Do not invent prompts I did not give you, and do not change the wording beyond tidying.
Finish with a list of anything that looks half-finished so I can decide whether to keep it.

7.2 Talking about AI credibly: your manager, your clients, your team

Sooner or later someone will ask. Your director wants to know whether "that AI thing" is safe before the board paper goes out. A client asks, half joking, whether a robot wrote their proposal. How you answer in that moment decides whether you are seen as the person who understands this or the person who is quietly taking risks.

The two failure modes are hype and fear. Hype ("it does everything, you have to try it") makes you sound like you have stopped checking. Fear (going quiet, or pretending you do not use it) makes a normal working practice look like something to hide. The credible position is in the middle, and it fits on one page.

Your one-page statement

Write a short document with four headings: what I use AI for, what I never put into it, how I check its output, and how I disclose it. First person, plain English, under a page. It is not a confession and not a policy for anyone else; it is your professional standard, written down so you can hand it over instead of improvising.

Under "what for", list real tasks: first drafts of routine emails, summaries of documents you have already read, notes into agendas, image and video briefs for Higgsfield and Canva as you did on Day 5. Under "what never", paste your rules from 7.1. Under "how I check", describe your Day 4 routine in two sentences. Under "how I disclose", say who you tell and when.

Answering "is this safe?" without flinching

Callum is marketing manager at Harbourline Dental in Brisbane. The practice manager noticed the patient newsletter arriving faster and asked whether patient information was going "into the internet". Callum did not defend the tool or apologise. He said what he does (drafts from public information and his own notes), what he never does (no patient names, records or appointment data in any AI tool), and what he would welcome (a written practice policy so everyone follows the same line). Two minutes later he had been asked to draft that policy.

That pattern (what I do, what I never do, what I would like) answers the fear directly with the confidentiality rules from Day 1, then moves the conversation forward instead of leaving it as a stand-off.

Proposing a small pilot

If your team is curious, do not propose "adopting AI". Propose one task, four weeks, and a before-and-after comparison. Pick something frequent, low-risk and already slightly painful: meeting notes into action lists, first drafts of internal updates, survey comments into themes.

Before you start, record how the task goes now: roughly how long it takes, how many rounds of correction, what people dislike about it. Agree the no-go list with the team first. After four weeks, measure the same things and add a column for what went wrong. Write it up on one page. You are not promising savings; you are offering to find out, which is far easier to say yes to.

Disclosure with clients

Mei runs a small practice at Maple Ridge Accounting in Vancouver. Her engagement letter says she uses AI tools to draft correspondence and summarise documents the client has already shared, that client financial data never goes into tools her firm has not approved, and that she remains responsible for every number and every word. She also asks whether the client has a policy of their own. One did, and it excluded their material from any third-party tool. Knowing that early saved an awkward conversation later.

Internally, tell your manager once and clearly rather than labelling every email. Externally, disclose when the client's data is involved, when they would reasonably want to know, or when your contract or professional body says so. When in doubt, "I use AI to draft; I check and I decide" is almost always true and almost always enough.

Draft a one-page "How I use AI at work" statement from my answers below.
My role: [YOUR ROLE AND THE KIND OF ORGANISATION YOU WORK IN].
What I use AI for: [LIST THE REAL TASKS].
What I never put into it: [PASTE YOUR NO-GO RULES].
How I check its output: [DESCRIBE YOUR CHECKS].
How I disclose: [WHO YOU TELL, AND WHEN].
Write it in the first person, in plain English, under 300 words, using those four headings. Keep the tone calm and factual: no praise for the tools and no apologising. Do not add uses, safeguards or claims I did not give you. End with one line saying when I will next review the statement.
Then, separately, list the questions a cautious manager or client might still ask after reading it. Do not answer them.

7.3 Staying current without the noise, and what's next

The tools you learned this week will change. Features will move, names will change, and something new will appear that everyone insists you must try. Your feed is already full of people announcing it. You do not have time to follow it all, and you do not need to. What you need is a small habit that catches the changes affecting your actual work and ignores the rest.

The fifteen-minute monthly check-in

Once a month, on the same Friday as your weekly review, add fifteen minutes and answer four questions in writing:

1. What changed in the tools I actually use? Not every tool; the two or three in your library.
2. What one new thing will I try next month, and on which task? One. Attached to a real task, or it will not happen.
3. Which prompt do I retire? Something you have not used, or something a newer entry does better.
4. What one thing is worth telling the team? Two sentences, sent the same day.

Four answers, each a line or two. If you cannot answer the first question, that is fine: nothing changed that matters to you, and you have your fifteen minutes back.

Where to look, and where not to

Three sources are enough. First, the tools' own update pages or release notes: every tool you use has one, and it is the only place where what changed is described without excitement. Second, your own notes: the Friday lines about what went wrong are the best signal you have about what to go looking for. Third, one trusted colleague who uses the same tools for similar work, with a standing agreement to tell each other about anything that changed a task.

Everything else (feeds, round-ups, people announcing that everything is different now) can wait. Apply one filter before you read anything: does this change a task I actually do? If the honest answer is no, close it.

Jordan leads a lettings team at Brightwater Property in Manchester. Their first monthly check-in took twelve minutes. The tool they use for long documents had changed how saved instructions work, so they rebuilt the tenant-letter assistant from Day 6, which took ten minutes. They retired a summarising prompt because the Day 4 decision-brief prompt covered it. The one thing for the team was a reminder that tenant names and addresses stay out of every tool. Nothing dramatic happened, which is the point.

Keeping the practice going

Skills fade when they stop being used on real work. Pick one task from this course to repeat each week until it is automatic; the CRAFT brief from Day 2 is the obvious first choice, because everything else depends on it. Refresh your voice file every few months, since the way you write shifts. Re-read your no-go rules whenever your employer's policy changes. And keep the worksheets: they are the record of what you tried and why.

What's next

Finish today's worksheet and mark Day 7 complete. Once all seven days are marked complete, your certificate of completion is available from the course home page. Then open the bonuses (the prompt library, the workplace safety checklist and policy template, and the which-tool-for-which-task cheat-sheet) and move anything useful into your own library, properly named. Your first Monday plan is where the real course starts.

I am doing my monthly fifteen-minute AI check-in. Here are my notes from the last four weeks: [PASTE YOUR FRIDAY NOTES].
Here are the names of the prompts in my library: [PASTE THE NAMES ONLY].
Here is what the tools I use say has changed: [PASTE OR SUMMARISE THEIR UPDATE NOTES, OR WRITE "NOTHING I NOTICED"].
Help me answer four questions in one short paragraph each:
1. What changed that affects a task I actually do?
2. What one new thing is worth trying next month, and on which task?
3. Which one prompt should I retire or merge, and why?
4. What one thing is worth telling my team, in two sentences?
Base everything on what I gave you. If my notes do not support an answer, say so rather than guessing.