Minutes 0–2: why models guess
A prompt is everything you give the model: your request, any text or files you paste in, and the conversation so far. The model has read a vast amount in general and nothing at all about your situation. It has never met your client, seen your last report or heard your manager’s preferences, unless you put them in the prompt.
When something is missing, the model does not stop and ask. It fills the gap with the most typical answer for a request like yours. That is why a vague prompt gets a confident answer fitted to an average case rather than yours. When the missing piece is a fact, it may invent one that sounds right, which is called a hallucination.
So the whole skill is leaving fewer gaps. The same habits work in ChatGPT, Claude, Gemini and any tool built on them. The vendors’ own guides from Anthropic, OpenAI and Google differ in detail, but all three start from clear instructions, relevant context, examples and a stated format.
Minutes 2–6: the five parts
- 1The task, and what it is forSay what to make and why. “A reply that keeps the client” gets a different answer from “a reply”.
- 2The materialPaste or attach what it should work from: the email thread, the notes, the figures, the policy.
- 3The finished resultWho will read it, how long it should be, what format, and what it must include.
- 4An example, when format mattersShow one good result or the layout you want, and say that it is an example.
- 5What to do when unsureAsk a question, mark the gap or say it does not know. Never fill a gap with a guess.
Each part closes a gap the model would otherwise fill on its own.
- The task and its purpose. “Summarize this report” has a hundred right answers. “Summarize this report so my director can decide by Friday whether to renew the contract” has a few. The purpose tells the model what to keep and what to drop.
- The material. If the answer depends on facts, give it the facts. The model cannot see your inbox, your shared drive or yesterday’s meeting unless you paste them in or your tool connects to them.
- The finished result. Name the reader, the length and the shape. “Five bullets my team can read on a phone” and “a one-page memo for the board” are different jobs.
- An example. Anthropic’s guide calls examples one of the most reliable ways to steer format and tone. Label it as an example, or the model may copy its content as well as its shape.
- What to do when unsure. Tell it how to handle gaps: “If a figure is missing, write ‘not stated’” or “Ask me up to three questions before you start.” Without this line, gaps get filled with guesses.
How much each task needs
Not every request needs all five. Two things decide how much to write: how much the task depends on what only you know, and how exact the result must be.
- Brainstorm names: Depends on what only you know 1/10, How exact the result must be 2/10
- Tidy a paragraph: Depends on what only you know 3/10, How exact the result must be 3/10
- Reply to a client: Depends on what only you know 8/10, How exact the result must be 4/10
- Meeting actions: Depends on what only you know 7/10, How exact the result must be 6/10
- Spreadsheet formula: Depends on what only you know 3/10, How exact the result must be 9/10
- Table from invoices: Depends on what only you know 9/10, How exact the result must be 9/10
Bottom left, a one-line request is fine: brainstorming names or tidying a paragraph you pasted. Moving right, the model needs your material: a client reply depends on the thread and the relationship. Moving up, it needs an exact format, so show one: a spreadsheet formula has to match your columns. Top right needs everything, and you should check the result before it goes anywhere: pulling figures from invoices into a table is precise work on your own documents.
Minutes 6–9: one task, before and after
An office manager has a proposal from a new cleaning contractor and has to give their manager a recommendation by Friday.
Summarize this proposal.
{{proposal}}I'm an office manager. We are choosing a cleaning
contractor, and my manager will decide on Friday
between this proposal and our current contract.
Proposal:
"""
{{proposal}}
"""
Our current contract:
"""
{{current_contract}}
"""
Write a note my manager can read in two minutes:
- First line: your recommendation, in one sentence.
- Then a short table comparing the two on monthly
price, what is included, contract length and
notice period.
- Then up to three risks or open questions, one
line each.
Use only these two documents. If a figure is
missing from either, write "not stated" in the
table instead of estimating it.The first prompt gets a fair summary of the proposal and nothing your manager can decide on. The second covers all five parts. The task and its purpose are in the first paragraph. The material is both documents, fenced with quotation marks so the model can tell them from the instructions. The finished result has a reader, a time limit and a shape. The table layout is the example. The last paragraph handles gaps, which matters most here: a model asked to compare prices can estimate a missing one, and an estimate in a table looks exactly like a fact.
Even the better answer needs a check. Compare each figure in the table with the documents before it goes to your manager. That takes a minute, and it is the habit that keeps an invented number out of a decision.
Minutes 9–12: fixing a weak answer
The first answer is a draft. How you respond to it matters as much as the prompt. Give feedback the way you would to a capable colleague: say what is wrong and what good looks like.
| Feedback that does little | Feedback that works |
|---|---|
| Make it better. | The recommendation is buried. Put it in the first line and cut the history of the company. |
| Too long. | Cut to 120 words. Keep the price, the start date and the notice period. |
| Sounds like AI. | Drop the opening compliment and the closing summary. Write the way this sample of mine reads: … |
| That’s wrong. | The notice period is 60 days, in section 4. Correct the table and check the other figures against the document. |
Three more habits help. If you are not sure what to ask for, end the prompt with “Before you start, ask me up to three questions about anything you need.” If a long chat has drifted, start a new one with an improved prompt rather than piling corrections on top of the old one. And when a prompt finally works, save it, so next week starts from the version that worked. The My Prompts keeps saved prompts with their version history.
Minutes 12–14: habits to drop
- Magic phrases and flattery. “You are a world-class expert” adds little on its own. A role helps when it carries real information. Anthropic’s guide notes that a one-sentence role focuses tone and behavior, so make that sentence say what the job is and who it is for.
- Shouting. Capital letters and “CRITICAL” or “you MUST” make some current models overdo the instruction, according to the same Anthropic guide. A plain instruction with its reason works better.
- “Think step by step” on every request. Many chat assistants now use reasoning models that work through a problem before answering. OpenAI’s reasoning best practices say asking them to think step by step is unnecessary and can sometimes hurt. Asking to see the working so you can check it is a different matter, and still worth doing.
- Asking for facts it cannot know. Your latest sales numbers, a client’s budget, last week’s news without search turned on. Supply the facts or ask it to mark them as gaps.
- Pasting confidential data into unapproved tools. Check which tools your company has approved for which data first. The risk checklist for AI at work has the short version.
- Trusting a prompt after one good answer. If a prompt will be reused, try it on five or ten real inputs first. The testing guide shows how without code.
Minute 15: what to do next
Take a prompt you used this week and rewrite it with the five parts. Then put it through the Grader. It checks the prompt against ten plain checks as you type, free and with no sign-up, and shows which gaps are left: no purpose, no material, no length, nothing about what to do when unsure. With your own API key, it also writes a rewrite for the model you use.
When you want more depth, the new-hire brief expands these five parts into eight with four more worked examples, and the ten checks explains each check the Grader runs.
Key takeaways
- Models fill every gap in a prompt with the most typical answer, and sometimes with an invented fact.
- A work prompt needs five things: the task and its purpose, the material, the finished result, an example when format matters, and what to do when unsure.
- The more a task depends on what only you know, the more material it needs. The more exact the result, the more an example helps.
- Treat the first answer as a draft and give specific feedback: what is wrong and what good looks like.
- Drop magic phrases, shouting and reflexive “think step by step”. Give real context instead.
- Check facts and figures before anything leaves your desk, and save the prompts that work.