Blog › Claude for Work › Part 2
How to Write a Prompt
That Doesn't Waste Your Time
Most people get mediocre results from AI because they write vague requests and expect detailed answers. Here's the prompting framework that actually works - with before-and-after examples for real work tasks.
The AI isn't the problem. The question is.
"Write me a project update." Claude produces something generic. You edit it for 20 minutes. You conclude AI isn't that useful.
But the problem isn't Claude. It's the prompt. "Write me a project update" is the equivalent of asking a new colleague to "do some work." Without context, constraints, or direction, the output will be a best guess at what you might want - and best guesses rarely hit.
Prompting is a skill. It takes about 20 minutes to learn the basics and it compounds every time you use it. Here's the framework that produces good results consistently, with examples you can copy and adapt for real work.
The four elements of a prompt that works
Role, context, task, format. In that order.
Role - tell it who it's being
Not literally - you don't need to say "pretend you're a PM." But giving Claude a frame helps it calibrate the output. "You're helping a project manager write a status update for an executive steering committee" produces a different result than no context at all. The frame shapes the vocabulary, the tone, and the assumed audience.
Context - give it what it needs to know
Claude knows nothing about your project, your organisation, or your situation unless you tell it. Background that feels obvious to you is not obvious to the model. The more relevant context you provide - project name, current status, audience, key issues, recent decisions - the more accurate and useful the output.
Task - be specific about what you want
"Write a summary" is vague. "Write a three-paragraph summary that covers what was decided, what's at risk, and what action is needed from the committee" is specific. Specific tasks produce specific outputs. Vague tasks produce vague outputs. Every time.
Format - tell it how to structure the output
Do you want bullet points or prose? A table or a list? Three paragraphs or ten sentences? Headers or no headers? If you don't specify, Claude will guess. It usually guesses well, but specifying takes five seconds and guarantees the output lands in a format you can use immediately.
Before and after - real examples
The same request, done badly and done well.
Example 1 - Status update
Weak prompt
"Write a project status update."
Strong prompt
"You're helping a project manager write a weekly status update for an executive steering committee. The project is a CRM system implementation for a financial services firm. We're in week 8 of 20. Key facts: go-live is on track, data migration is 2 weeks behind, vendor delivered UAT environment last Tuesday. The committee needs to know status, risks, and what decisions they need to make. Write a 200-word update in three sections: Overall Status, Key Risks, and Decisions Required. Use plain language, no jargon."
Example 2 - Summarising a document
Weak prompt
"Summarise this document." [pastes document]
Strong prompt
"I'm a project manager briefing an executive who hasn't read this document and has 3 minutes. Summarise it in 5 bullet points. Each bullet should be one sentence. Focus only on decisions made, risks identified, and next actions required. Ignore background and methodology sections." [pastes document]
The pattern: the strong prompts aren't longer for the sake of it. Every additional sentence adds a constraint that removes a decision Claude would otherwise have to make by guessing. Fewer guesses, better output.
Five prompting habits that change results immediately
Pick any two of these and your output quality will improve today.
State the audience
"Written for an executive who doesn't have technical background" produces different language than "written for the delivery team." Claude adjusts vocabulary, assumed knowledge, and depth based on who's reading.
Give a word or length limit
"In under 150 words" or "in three sentences" forces Claude to prioritise. Without a limit, it'll err toward completeness. With a limit, it'll edit itself. You get the essence, not the padding.
Say what to leave out
"Don't include background or methodology" or "don't offer caveats" tells Claude what to skip. Negative constraints are as powerful as positive ones and often more specific.
Ask for a draft, not a final
"Give me a first draft I can edit" removes pressure for perfection and often produces a better starting point than asking for something "polished." It also signals you'll be reviewing it.
Ask it to ask you questions first
For complex tasks: "Before you write anything, ask me the three questions you'd need answered to do this well." This surfaces what context is missing before Claude starts guessing.
Iterate, don't restart
If the first output is 70% right, tell it what to fix: "Good start. Shorten the second paragraph, make the tone less formal, and add a clear next action at the end." You don't need a new prompt. You need a refinement.
Prompt templates you can use right now
Copy, fill in the brackets, send.
Meeting summary
"Summarise the following meeting notes into: (1) decisions made, (2) actions with owners and due dates, (3) open issues still to resolve. Use a table for actions. Keep each section to five points maximum. Here are the notes: [paste notes]"
Stakeholder email
"Draft an email to [role] about [topic]. The key message is [one sentence]. The tone should be [professional/warm/direct]. It should be under 100 words. Don't use jargon. End with a clear ask: [what you need from them]."
Risk assessment
"I'm a project manager working on [project type]. Identify the top 5 risks for this type of project. For each risk, give: a one-line description, a probability rating (high/medium/low), an impact rating (high/medium/low), and one mitigation action. Present as a table."
Explain something to a non-technical audience
"Explain [technical concept] to a senior executive who is not technical and has 2 minutes. Use plain English. No acronyms unless you explain them. Use one real-world analogy. Keep it under 100 words."
The one thing most people skip
Context is the multiplier.
The single biggest gap between people who find Claude useful and people who don't is context. People who get good results dump relevant context into the prompt without worrying whether it's "too much." People who get poor results write brief requests and expect the model to fill the gaps.
Claude's context window is large. You can paste in a long document, a string of meeting notes, a policy, a template - anything relevant - and it'll process all of it. The cost of including too much context is minimal. The cost of including too little is a generic, unhelpful response.
When in doubt, add more context. The prompt you're embarrassed by for being "too long" is usually the one that produces the best output.
Try it now
Take one of the templates above, fill in your details, and paste it into Claude's free tier. The difference from a vague request is immediate.
Open Claude at claude.ai →Claude for Work - series
Part 1 - What Claude Actually Is
Part 2 - How to Write a Prompt That Doesn't Waste Your Time (you're here)
Part 3 - How to Use Claude to Draft a Project Plan in 10 Minutes (coming soon)
Part 4 - How to Use Claude for Stakeholder Updates (coming soon)
Part 5 - Is Claude Safe for Work Documents? (coming soon)