← Learn
Playbook · 10 steps · checked 4 October 2026 · 0 adopted

Prompting

At most ten steps, in order of how much they help. Each step is from Anthropic's own pages, linked. Re-checked against those pages every week; anything that can't be confirmed is marked unverified rather than removed.

  1. 01

    Write it so a new colleague could follow it

    Treat Claude as a brilliant new employee who does not know your norms. Say exactly what you want, including format and limits, and use numbered steps when order matters. If you want more than the basics, ask for it. Test: would a colleague with no background be confused by your prompt? Then Claude will be too. Start simple and add detail only when the output falls short. (best practices, prompting blog post)

  2. 02

    Say why, not just what

    Give the reason behind a rule. "Never use ellipses" works less well than "this will be read aloud by a text-to-speech engine, so never use ellipses". Claude generalises from the reason, so it also handles cases your rule did not name. (best practices)

  3. 03

    Show an example or a few

    Examples are one of the most reliable ways to steer format, tone and structure. Make them close to your real task and varied enough that Claude does not copy an accident of one example, and wrap them in <example> tags. The docs suggest 3 to 5 for best results; the blog suggests starting with one and adding more only if the output still misses. You can ask Claude to check your examples or write more. (best practices, prompting blog post)

  4. 04

    Long documents first, question last, quotes before answers

    For inputs of roughly 20k tokens or more, put the documents at the top and your question and instructions at the end; Anthropic reports this can improve quality by up to 30 percent in tests. Wrap each document in <document> tags with its <source>. Then ask Claude to pull out the relevant quotes first and answer from those. In a Project, files in the project knowledge are used across all chats in that project; a chat's own content is not shared with other chats unless you add it there. (best practices, Projects help)

  5. 05

    Separate the parts with XML tags

    When a prompt mixes instructions, background, examples and input, wrap each in its own tag, such as <instructions>, <context> and <input>. Use the same clear tag names every time, and nest them when the content nests. This stops Claude mistaking your data for your instructions. (best practices)

  6. 06

    Describe the output you want, in the style you want

    Say what to do rather than what not to do ("write flowing prose paragraphs", not "no markdown"). Your prompt's own style rubs off, so a prompt without markdown tends to get less markdown back. For JSON or labels in the API, use Structured Outputs or a tool with an enum field. Prefilling the assistant's reply is not supported from the Claude 4.6 models onwards (it returns an error), so ask directly, for example "Respond directly without preamble". (best practices, moving off prefill)

  7. 07

    Let Claude think, and ask it to check its work

    Current models decide for themselves when and how much to think (adaptive thinking); in the API, the effort setting (low to max) is the main control for how hard it works. A general instruction such as "think thoroughly" often beats a hand-written step-by-step plan. Add "Before you finish, verify your answer against [your criteria]" for coding and maths. Model notes: on Claude Opus 5, remove such checking instructions, since it checks its own work well and they can cause over-checking; on Claude Fable 5.1, Fable 5, Opus 5.5, Opus 5 and Sonnet 5.5, rely on built-in thinking rather than asking it to write reasoning in <thinking> tags, which may be declined. Claude Opus 5.5 defaults to medium effort, most others to high. (best practices, effort)

  8. 08

    Allow "I don't know" and keep it to your sources

    Tell Claude it may say it lacks enough information; this alone can sharply cut made-up answers. When accuracy matters, tell it to use only the documents you gave, cite a quote for each claim, and drop any claim it cannot back with a quote. Still check anything high-stakes yourself. (reduce hallucinations)

  9. 09

    Put standing instructions where they persist

    A single sentence of role in the system prompt focuses tone and behaviour ("You are a helpful coding assistant specialising in Python"). In the API that is the system field; in a Project, use "Set project instructions", which apply to every chat in that project. For Claude Code, see [learn/claude-code.md](claude-code.md). (best practices, Projects help)

  10. 10

    Decide what "good" means, then test against it

    Before tuning, Anthropic suggests having clear success criteria, a way to test against them, and a first draft. Write specific, measurable criteria, collect test cases that look like real use plus edge cases (missing, overly long or irrelevant input), and grade automatically where you can; many cases with rough automatic grading beat a few hand-graded ones. Change one thing at a time and re-run. In this repo, /prompt test can stress-test a prompt. No draft yet? The overview points to the metaprompt notebook in the Claude Cookbook; the old experimental prompt generator and improver APIs were retired on 17 August 2026. (overview, define success and build evals, release notes)

Picked from around the web, with a one-line takeaway. The originals are by their creators.

  1. instagram

    Nate, a designer at Anthropic, used to start a presentation by staring at a blank slide ↗

    Claude (@claudeai) · still works (checked 6 Oct 2026)

    Shows a way to start a deck: talk the idea into a voice memo, then give Claude the transcript and ask it to organize your exact words into slides.