GuideClaude and ChatGPTBeginner

Write prompts that get better answers from Claude and ChatGPT

Eight habits that Anthropic and OpenAI both recommend in their own prompting guidance, with before-and-after prompts you can adapt for everyday work.

Curated by Javier Díaz · Written with Claude from official documentation · Checked 18 October 2026 · 15 min

Anthropic and OpenAI both publish advice on how to prompt their models. Read side by side, the two sets of guidance agree far more than they differ: say exactly what you want, explain why, keep your instructions apart from your material, show an example, ask for a specific format, and refine as you go.

This guide gathers those shared techniques for people who use Claude or ChatGPT in the chat app for ordinary work: drafting, summarising, reviewing documents, planning. Much of the official material is written for developers using the API, but the ideas carry straight over to a chat box. Where the two companies' advice differs, the guide says so.

You don't need every technique for every request. A quick question deserves a quick prompt. Reach for these when an answer comes back generic, in the wrong shape or off target, or when the task matters enough that you would brief a colleague properly.

1. 01Say exactly what you want, as you would to a new colleague

   Both companies open with the same point: a clear, specific request beats a vague one. Anthropic suggests picturing a capable new starter who knows nothing about how your team works. OpenAI suggests writing as though you were asking another person, who would also struggle with a one-line brief.

   A good check before you send: would a colleague with no background on the task understand what you want? If not, the model won't either. Name the task, the audience, the length and what counts as finished. Swap loose words for measurable ones: "short" becomes "under 150 words", "a few ideas" becomes "five ideas".

   Before:

   Write something about our new expenses policy.
   

   After:

   Write a 200-word item for the staff newsletter about the new expenses policy, which starts on 1 November. Most readers skim, so lead with the one change that affects everyone: receipts must be uploaded within 14 days. Close by saying that exceptions go to the finance team.
   

2. 02Explain the context and the reason behind the request

   Anthropic's guide notes that explaining why an instruction matters helps Claude aim its answer, and lets it apply the rule sensibly to cases you didn't spell out. OpenAI's ChatGPT guidance likewise asks you to give enough context for the model to understand what you are after.

   Useful context is usually short: who the output is for, what it will be used for, what you have already tried, and any constraints you are working within.

   Before:

   Make this slide text shorter.
   

   After:

   Make this slide text shorter. I have a ten-minute slot at the board meeting and they will read the slide while I talk, so each bullet must make sense at a glance. Keep every figure; cut the explanations, because I will say those out loud.
   

3. 03Keep your instructions separate from your material

   When you paste in an email, a report or a page of notes, mark clearly where it starts and ends. OpenAI's help centre suggests simple delimiters such as ### or triple quotes. Both companies' developer guides recommend XML-style tags, which Anthropic says help Claude read prompts that mix instructions, background and examples without confusing one for another. The tag names aren't special; make them descriptive and use them consistently.

   Below are a customer's complaint and our refund policy. Draft a reply that applies the policy fairly and offers the customer a clear next step.
   
   <complaint>
   [paste the email here]
   </complaint>
   
   <policy>
   [paste the policy here]
   </policy>
   

   The two companies differ slightly on order. For long documents, Anthropic recommends putting the material first and your question at the end, and reports better answers that way in its tests. OpenAI's help centre suggests putting the instruction at the start. With a short paste, either works. With a long one, a sensible compromise is a one-line instruction at the top and the full question again at the bottom.

4. 04Show an example of what good looks like

   Anthropic describes examples as one of the most reliable ways to steer format, tone and structure, and OpenAI's help centre makes the same point: models do better when you show the format rather than only describe it. Anthropic suggests three to five examples, varied enough that the model doesn't copy incidental details, and wrapped in tags so they aren't mistaken for instructions.

   Sort each piece of customer feedback into one of: Billing, Bug, Feature request, Praise. Reply with the label only.
   
   <examples>
   <example>
   Feedback: I was charged twice in September.
   Label: Billing
   </example>
   <example>
   Feedback: The export button does nothing in Firefox.
   Label: Bug
   </example>
   <example>
   Feedback: A dark mode for the dashboard would be lovely.
   Label: Feature request
   </example>
   </examples>
   
   Feedback: The new sign-up screens are much clearer, thank you.
   

   One caution from OpenAI's guidance on reasoning models: try without examples first, because those models often don't need them, and if you do add examples, make sure they agree with your written instructions. An example that contradicts the brief can pull the answer the wrong way.

5. 05Ask for the format, length and tone you want

   If the shape of the answer matters, say so. Both companies advise stating length and format explicitly, and both recommend saying what to do rather than only what to avoid: "write in short paragraphs" works better than "don't use bullet points". For tone, OpenAI's ChatGPT guidance suggests plain descriptive words such as formal, friendly or serious. Anthropic adds that the style of your prompt can rub off on the reply, so a prompt full of headings and bullets tends to get the same back.

   Before:

   Reply to this client. Don't make it too long or too stiff.
   

   After:

   Reply to this client in three short paragraphs, friendly but professional, with no headings or bullet points. End with one clear question asking them to confirm the delivery date.
   

   For comparisons, name the structure outright, for example: "Give me a table with the columns Option, Cost, Main risk and Your recommendation."

6. 06Give the model a role when expertise or perspective matters

   Anthropic says that even one sentence setting a role focuses Claude's tone and behaviour. OpenAI's prompt guide suggests that a set of instructions begins with the model's identity: who it is meant to be and what it is for. Developers put this in a system or developer message; in the chat app, you can open your message with it.

   You are an experienced editor of business reports. I'm a project manager, not a writer. Read the executive summary below and tell me the three changes that would most improve it for a busy director, with a rewritten example of each.
   
   <summary>
   [paste the summary here]
   </summary>
   

   If you find yourself giving the same role or background in every chat, both apps let you save it once. In Claude, it goes in the Instructions for Claude section of Settings. In ChatGPT, use custom instructions, found under Personalization in Settings on the web. A role adds a perspective; it doesn't replace the task and context from the earlier steps, so use it alongside them.

7. 07On hard problems, set the goal and ask it to check its work

   Current models from both companies can reason before they answer, and the guidance on prompting them has changed. OpenAI's reasoning guide says telling those models to think step by step is unnecessary and can even make results worse. Anthropic's guide says a general request to think carefully often does better than a step-by-step plan you write yourself, and notes that some of its newest models keep their reasoning internal and may decline a request to write it all out.

   What helps instead is a clear end goal and firm constraints, which OpenAI recommends, plus a request to check the answer against them, which Anthropic recommends.

   We need to cut £40,000 from next year's marketing budget (pasted below). Propose two options. Each must keep the spring trade show and must not cut any single channel by more than half. Before you answer, check both options against those two rules and the £40,000 target, and show the totals.
   

   For a big job with several distinct parts, both companies suggest breaking it into smaller requests in sequence: for example, outline first, then draft, then review the draft against your criteria, then revise.

8. 08Treat the first answer as a draft and refine it

   Both companies describe prompting as iterative. OpenAI's ChatGPT guidance suggests starting with a first attempt, reading the result, then adjusting the wording, adding context or simplifying the request. Treat it as a conversation rather than a single shot.

   Specific feedback works best. Point to what is wrong and say what you want instead:

   Closer. Keep the structure, but the second paragraph assumes the reader knows what our CRM is. Add one sentence explaining it, and drop the final paragraph.
   

   If a chat has drifted a long way from what you need, it can be quicker to open a new one with an improved first prompt that includes everything you learned. When a prompt works well, keep a copy so you can reuse it next time.

Written by Claude from the official pages above and checked against them on 18 October 2026. Products change; if a step no longer matches what you see, the sources have the current version. How we work