AI prompts for work that actually hold up.
Not another list of 100 prompts. Seven briefs that cover most of an office week, the pattern behind them, and how to adapt each one so it works for your job rather than an imaginary average one.
Downloaded prompt lists disappoint for a predictable reason: a prompt written for nobody in particular produces output for nobody in particular. The assistant is not weak - the brief is empty. What holds up at work is a small set of prompts built like proper briefs, refined each time you use them.
What does a good AI prompt for work look like?
The prompts that hold up at work are the handful you keep and refine, not a downloaded list of a hundred. Each one below is a ready-made brief for a specific job - summarising a long document, breaking a blank page, pressure-testing a decision, proofreading without a rewrite. Adapt them rather than collecting more.
The method these are built on - how to brief an assistant so it gives you something usable - is set out in how to use AI at work. Go there for the method. This page is the library: the prompts themselves, ready to paste, plus pointers to the task guides where a job has an article of its own. They work in ChatGPT, Copilot, Gemini and Claude alike, because a good brief does the work, not the tool. Square brackets mark the parts you replace.
The document summariser.
Twenty pages, five bullets.
"Summarise this document for a [busy manager / new starter / client] who will spend 60 seconds on it. Five bullets maximum: lead with anything that requires a decision or action, then the facts that change what we do, then context. Flag anything ambiguous rather than smoothing it over. Document: [paste]"
Naming the reader changes the summary more than any other word. A summary "for a busy manager" surfaces decisions; the same document summarised "for a new starter" surfaces definitions. Say who it is for, every time.
How do you get a first draft out of bullet points?
First draft from bullets.
"Draft a [status update / proposal section / announcement] from the bullets below. Audience: [who]. Length: [X] words. Tone: plain, confident, no exclamation marks, no words like 'exciting' or 'delighted'. Do not add facts I have not given you. Here is an example of the style I want: [paste a past piece you were happy with]. Bullets: [paste]"
The pasted example is the single highest-value habit in this article. One good past example outperforms a paragraph of tone description, because the assistant can copy what it can see.
The thinking partner.
Argue against me.
"I am about to [decision or plan]. Here is my reasoning: [paste]. Play the sceptical colleague: give me the three strongest objections, what each would mean in practice, and one question I should be able to answer before going ahead. Do not soften the objections."
Assistants default to agreeable. Instructed to disagree, they are genuinely useful before a meeting where you expect pushback - not because they are right, but because they are fast at generating the objections you will face.
How do you proofread without losing your voice?
Fix errors, keep me.
"Proofread the text below. Fix spelling, grammar and punctuation only - do not rephrase sentences, change vocabulary or alter the tone. List every change you made as a numbered list (original → corrected), then give me the corrected text. Then list separately any sentences that are unclear, without rewriting them. Text: [paste]"
Without the restriction, every assistant "improves" your writing into the same beige corporate register. Splitting fixes from suggestions keeps you in charge of the rewriting. The numbered list of changes is what makes the restriction checkable: assistants are unreliable at obeying a "do not rewrite" instruction, and a quietly reworded sentence buried in a returned block of text is invisible. Listed as original and corrected, you can scan the lot in seconds and reject the ones that were not errors.
Which prompts live in the task guides?
Three jobs have enough rules of their own to need a full article rather than a single prompt, so they are not duplicated here. For email, the move that matters is pinning the tone to your own past messages rather than the loudest voice in the thread; the untangling prompt, the polite chase, the diplomatic no and the messages AI should never write are in using AI for work emails. For meetings, the line that earns its place in any minutes brief is "do not guess owners", and the brief itself, the capture options, the recording etiquette and the checks before circulating are in AI for meetings. For spreadsheets, asking how a formula could silently give a wrong answer is the difference between a party trick and a working tool; the explainer prompt, plus formula writing, data cleaning and first-pass analysis, are in AI for spreadsheets.
How do you adapt these prompts to your own job?
Treat each prompt as a template with three dials: the audience, the format, and the quality line. When an output disappoints, resist the urge to start again - tell the assistant what was wrong ("too formal", "you invented the June figure", "half the length") and let it revise. Then fold that correction into your saved prompt so next time starts better. A prompt refined over five uses becomes precisely fitted to your job, which is why a short library of your own will beat anyone's list of a hundred. And remember the boundary that applies to all of them: raw material goes in only if it is safe to share - the rules are here.
Frequently asked questions
What makes a good AI prompt for work?
Four parts: context (who you are and who the output is for), the task and format, your raw material pasted in, and one line on what good looks like. Prompts fail when they skip context and standards, not because they lack magic words.
Do the same prompts work in ChatGPT, Copilot, Gemini and Claude?
Yes. A well-structured brief works across all mainstream assistants because it supplies context, material and standards rather than relying on tool-specific tricks. Output style differs slightly between tools, so keep your one-line quality standards in every prompt.
Should I save my work prompts?
Yes. Keep a notes file of the five to ten prompts you reuse, and refine the wording each time an output disappoints. A refined prompt library is more valuable than any downloaded list because it encodes your standards and your job's context.
Why do prompts from big prompt lists give poor results?
Most published prompts are generic: they contain no context about you, your reader or your standards, so the assistant fills the gaps with averages. Adapting one of your own past outputs as a style example beats any generic prompt.
The library, in one line each.
- Summarise for a named reader: decisions first, then the facts that change what you do.
- Break the blank page: bullets in, plus a past piece you liked as the style example.
- Pressure-test a decision: ask for the three strongest objections, unsoftened.
- Proofread without a rewrite: errors only, listed as original and corrected.
- For email, meetings and spreadsheets, use the task guides - those prompts live there.
- Refine, don't restart: correct the output, then fold the correction into your saved prompt.
Put AI to work in your own job.
AI at Work is Clerq's practical, hype-free guide to getting real work done with AI - written for people with inboxes, deadlines and meetings, not developers. PDF and EPUB, launching soon.