How to write AI prompts that work

If you have ever asked an AI assistant for something and received a bland, generic or slightly wrong answer, the likely cause is the request. An assistant cannot see your inbox, your client or your standards. It only sees what you type, so it fills every gap with an average guess.

This guide gives you a method that works in Claude, ChatGPT, Copilot or any other chat assistant as of September 2026. It has no tricks or magic phrases. It is a way of briefing the tool the way you would brief a capable new colleague on their first day.

What a prompt is doing

A prompt is the instruction and material you give the assistant. The assistant predicts a useful response from that text, so the quality of the response tracks the quality of what you supplied. Three things change results the most:

1. Context. Who is this for, and what is the situation?
2. A clear task. What exactly should come back?
3. A picture of good. What length, format and tone count as done?

If you leave one out, the tool guesses. Sometimes the guess is fine. For work that matters, don’t leave it to chance.

Step 1: State the job and the reader

Start with one sentence about the task and one about who will read the result.

Weak:

Write an email about the delay.

Better:

Write an email to a client, Priya, telling her the software rollout will slip from October 15 to October 29 because a vendor missed a delivery. She is senior, cares about dates and doesn’t like surprises.

The second version tells the assistant what happened, who is reading and what she cares about. It can now choose a tone and a structure on its own.

Step 2: Give it the real material

Assistants write better when they can work from your material instead of inventing it. Paste in the relevant email, notes, numbers or policy text, and label each piece so it isn’t confused with your instructions.

Here is the last email thread with Priya:

[paste thread]

Here are the facts you can use: new date October 29; cause is a late shipment from our vendor; we are adding a weekly checkpoint call at no cost. Don’t add any facts that aren’t listed.

That last sentence is worth copying into many prompts. It tells the tool to stay inside what you provided, which cuts down on invented details. Before you paste anything, make sure it is allowed. See our guide to a company AI policy and the question is it safe to put company data in ChatGPT.

Step 3: Say what the output should look like

Describe the format the way you would to a person: length, structure, tone, and anything to avoid.

Format: a subject line plus three short paragraphs, under 150 words total. Tone: direct and warm, no apologies beyond one sentence. Avoid corporate phrases like “circle back” and “going forward.”

If you have a good past example, paste it and say so: “Match the style of this earlier email.” Showing beats describing.

Step 4: Put it together with a reusable template

Here is a template that combines the first three steps. Copy it into a note and fill in the brackets.

Task: [what you want done, in one sentence]

Audience and situation: [who will read it, what they care about, what just happened]

Material to use:
[paste the source text, numbers or notes]

Rules: Use only the facts above. If something is missing, ask me instead of guessing.

Output: [format, length, tone]

The “ask me instead of guessing” line is a strong default. Many assistants will then come back with one or two clarifying questions, and answering them takes less time than fixing a wrong draft.

Step 5: Ask for a plan on bigger jobs

For anything with several parts, such as a report, an analysis or a proposal, ask for an outline first.

Before writing anything, give me a five-bullet outline of how you would structure this and list any information you’re missing. I’ll confirm, then you write it.

You catch a wrong direction in ten seconds instead of after a full draft. This is also how you avoid getting a long answer that misses the point.

Step 6: Iterate with specific feedback

The first response is a draft. Improve it with instructions that name the problem.

• Weak feedback: “Make it better.”
• Better feedback: “The second paragraph is too defensive. Cut it to one sentence and lead with the new date.”
• Better still: “Keep everything, but rewrite the opening so the new date appears in the first sentence.”

Small, precise corrections beat starting over. If a long conversation starts to wander, open a fresh chat and paste in the current version plus a short brief. Long chats can lose track of early instructions because the assistant only holds so much text in view at once (this limit is called the context window).

Other useful follow-ups:

• “Give me three versions: shorter, warmer and more formal.”
• “What did you assume that I didn’t tell you?”
• “What would a skeptical reader push back on here?”

Step 7: Use roles and examples sparingly and honestly

You will see advice to open every prompt with “You are the world’s best copywriter.” Roles can help when they add real context, such as “You are reviewing this as a finance controller who must approve the spend.” A flattering title without context does little.

Examples are more powerful. Two examples of the output you want (“here is a good summary, and here is a bad one and why”) often fix a stubborn tone or format problem quickly.

How to check the output

A polished answer isn’t necessarily a correct one. Run these checks before you use anything.

1. Facts against source. Compare every date, number, name and claim to your material. Ask: “List each factual claim in your draft and point to where in my material it comes from.”
2. Invented details. Look for specifics you didn’t supply, such as a statistic, a quote, a policy name. Delete or verify them. This problem is called hallucination; see how to stop AI hallucinations.
3. Instruction check. Did it follow the length, tone and format? Ask: “Which of my instructions did you not follow?”
4. Reader test. Read it as the recipient. Would it land the way you want?
5. Sensitive content. Confirm nothing confidential slipped in that shouldn’t leave your company.

For low-stakes drafts one pass may be enough. For anything going to a client, regulator or board, read every line yourself.

Worked example: from weak to strong

Say you want a summary of survey results for your leadership team.

Weak prompt:

Summarize this survey.

Strong prompt:

Task: Summarize the attached employee survey results for the leadership team.

Audience and situation: Six executives who will read this in 2 minutes before a Monday meeting. They want to know what changed since last year and what needs a decision.

Material to use:
[paste results table and last year's table]

Rules: Use only the numbers provided. Show change as percentage points. If a number is missing, say so.

Output: A headline sentence, then five bullets ordered by size of change, then one line titled "Decision needed". Under 200 words.

The second prompt takes an extra minute to write and typically saves several rounds of correction. Better still, save it.

Build a small prompt library

Once a prompt works, keep it. Create a note or document with a section for each recurring job: client updates, meeting follow-ups, weekly summaries, first-draft replies. Store the template, the date you last used it, and a line on what to check afterwards. Teams that share such a library avoid everyone rediscovering the same fixes. If you use ChatGPT, a saved assistant such as a custom GPT can hold the instructions for you; in other tools, look for saved instructions or projects.

Related reading: how to write better AI prompts, the glossary entries for prompt and prompt engineering.

Common mistakes

• Asking for the finished product with no context. Give it the reader and the material.
• Piling on rules. Ten conflicting instructions confuse the result. Prioritize three or four.
• Never saying no. Tell it what to avoid as well as what to do.
• Trusting fluent output. Confidence is not accuracy.
• Starting over every time. Edit the prompt, and save the version that worked.

What Anthropic’s prompting guidance says

Anthropic’s documentation for Claude gives advice that matches the method above.

• Anthropic’s documentation says to think of Claude as a brilliant but new employee who lacks context on your norms and workflows.
• Anthropic’s Claude documentation suggests showing your prompt to a colleague with minimal context on the task and asking them to follow it.
• Anthropic says examples are one of the most reliable ways to steer Claude’s output format, tone, and structure.
• Anthropic advises telling Claude what to do instead of what not to do.
• In the arXiv paper on GPT-3, the authors report that scaling up language models greatly improves performance with only a few examples.

Keep going

Once your prompts are reliable, apply them to real work: meeting notes that turn into action items, summarizing PDFs and contracts or cleaning spreadsheets with AI.

In our two-day class we spend the first morning on a framework called CRAFT (Context, Request, Alternatives, Feedback, Transfer). The whole room takes the same board update memo from a naive prompt to a strong one, live, and you can reuse the framework afterward.

Sources

• Anthropic, Prompting best practices (Claude Platform Docs): “Think of Claude as a brilliant but new employee who lacks context on your norms and workflows”
• Anthropic, Prompting best practices (Claude Platform Docs): “Show your prompt to a colleague with minimal context on the task”
• Anthropic, Prompting best practices (Claude Platform Docs): “Examples are one of the most reliable ways to steer”
• Anthropic, Prompting best practices (Claude Platform Docs): “Tell Claude what to do instead of what not to do”
• arXiv, Language Models are Few-Shot Learners: “scaling up language models greatly improves task-agnostic, few-shot performance”