Short answer. A good prompt says who the model should be, gives the facts it cannot know, names one specific task, sets limits, and describes the shape of the answer you want. When the first answer misses, change one part of the prompt at a time rather than starting again.
Prompt collections
7 promptsPrompts for AI agents and assistantsShorter answers, lower token use, no loops, no needless disclaimers, progress summaries, asking before acting, and hard limits. Open the collection →Related: AI writing tools, compared and reviewed.
Building a prompt, part by part
Start from a one-line request and add each part in turn. Watch the prompt grow, and read what each addition changes in the answer.
Write an email about the price increase.You are an account manager at a small design studio.Our monthly fee rises from £450 to £495 on 1 September, our first increase in three years. The client is a long-standing bakery chain.Write the email telling the client about the change and thanking them for their loyalty.Under 120 words. No apology for the increase. Mention that service levels stay the same.Return a subject line, then the email body, with no commentary.
What this adds The shape of the reply. You can paste the result straight into your email client.
Weak and strong prompts compared
A weak prompt names a topic. A strong prompt names a job. These four everyday requests show the difference; the right-hand column explains what the stronger version fixes.
| Task | Weak | Strong | What changed |
|---|---|---|---|
| Summarise | Summarise this report. | Summarise the report below for a busy finance director in 5 bullet points. Lead with the decision they need to make. Quote every figure exactly. | Names the reader, the length, what comes first and a rule against rounding numbers. |
| Rewrite | Make this better. | Rewrite this paragraph so a new customer understands it on first reading. Keep every fact. Use sentences under 20 words. Show only the new version. | “Better” becomes a test you can check: clear to a newcomer, no facts lost, short sentences. |
| Brainstorm | Ideas for a team event. | Suggest 8 team events for 12 people in Manchester, under £40 each, suitable for people who do not drink. One line per idea with the rough cost. | Number, place, budget and one firm limit turn a vague list into options you can actually book. |
| Explain | Explain the Russian cases. | Explain what a grammatical case is to someone who only speaks English, using one example sentence. Then list the six Russian cases with one line each. | Sets the reader’s starting point and the order: concept first, then detail. |
Five prompt patterns
These patterns work with any assistant, including ChatGPT, Claude and Gemini. Each solves one recurring problem; combine them freely.
Show an exampleUse it when the format matters more than the content Write product descriptions in this style:
Example: “Oak desk. 120 cm. Seats two. Ships flat.”
Now write one for: a walnut bookshelf. One good example teaches a style faster than a paragraph describing it. This is called few-shot prompting.
Separate the materialUse it when you paste in long text Summarise the email between the ### marks.
###
[paste email]
### Clear markers stop the model confusing your instructions with the text it should work on.
Ask it to ask youUse it when you are not sure what details it needs I need a cover letter. Before writing, ask me up to five questions about the job and my experience. The model gathers the context itself, so the first draft is already specific.
Think step by stepUse it when the task involves reasoning or numbers Work out the total cost step by step, showing each calculation, then give the final figure on its own line. Asking for the working makes errors visible, and you can check each line.
Draft, then critiqueUse it when quality matters more than speed Write the announcement. Then list its three weakest points and write an improved version. A second pass catches vague phrases and missing facts that the first pass let through.
Asking for a tone
“Make it professional” means different things to different readers. Describe the relationship and the feeling instead. Here is one request, declining a meeting, with three tone instructions and the opening line each one produces.
Formal Decline the meeting invitation from a senior client I have not met. Courteous and formal, no first names. “Dear Ms Patel, thank you for your kind invitation to meet on Thursday.”
Neutral Decline the meeting from a colleague in another team. Clear and friendly, straight to the point. “Hi Sam, thanks for the invite. I can’t make Thursday, but could we do Friday at 10?”
Warm Decline the coffee meeting from a former colleague I like. Warm and personal, suggest another date. “Lovely to hear from you, Jo! Thursday is packed, but I’d really like to catch up.”
Fixing a prompt that is not working
- 1Name what is wrongToo long, too vague, wrong tone, missing a fact? Write it down in one phrase before changing anything.
- 2Find the missing partToo generic usually means missing context; wrong shape means missing format; too long means missing constraints.
- 3Change only that partAdd or rewrite one block of the prompt, keep the rest the same, and run it again.
- 4Tell it what to do, not only what to avoid“Use plain words a 12-year-old knows” works better than “do not use jargon”.
- 5Save the version that worksKeep the finished prompt in a note so you can reuse it with new details next time.