Writing & chat · guidance updated 22 Nov 2026

Claude prompt guide

Current models include Fable 5.2, Mythos 5.2, Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5.5, Sonnet 5, Sonnet 4.6, Haiku 5.5, Haiku 4.5. Techniques apply primarily to Fable/Mythos 5.2, 5.1, 5. Effort calibration, approval flows, subagent delegation in Fable/Mythos 5.2.

Free. We don't store what you type.

How to prompt Claude

  1. Read your model's dedicated prompting page first before applying general techniques.
  2. Establish success criteria and empirical evaluation methods before prompt engineering begins.
  3. Determine whether prompt engineering or model selection best solves your problem.
  4. Be explicit about scope; ask for 'fully featured' or 'go beyond basics' if needed.
  5. Explain why each constraint matters; Claude generalizes from your reasoning.
  6. Wrap distinct content in XML tags: <context>, <instructions>, <examples>, <documents>.
  7. Place long documents (20k+ tokens) above your query for improved performance.
  8. Use <document> tags with <source> and <document_content> subtags for multiple inputs.
  9. Give 3–5 diverse, relevant examples in <examples> tags when tone or format must match.
  10. Say 'do X' rather than 'suggest X' when you want the model to produce output.
  11. Describe desired format positively ('flowing prose') not as bans.
  12. Use role-setting (one sentence) only when domain expertise or tone matters.

The shape of a good prompt

[One-line role, optional]
<context>[background, audience, why it matters]</context>
<documents>
  <document index="n">
    <source>[name/identifier]</source>
    <document_content>[material]</document_content>
  </document>
</documents>
<instructions>[numbered steps if order matters]</instructions>
<examples><example>[diverse, relevant sample]</example></examples>
<format>[length, structure, tone—positive description]</format>
[Query at end]

Avoid

  • Skipping model-specific guidance; each model has its own dedicated prompting page.
  • Starting prompt engineering without defined success criteria and evaluation design.
  • Using prompt engineering to solve problems better addressed by model selection.
  • Vague requests that rely on the model inferring ambition level.
  • Bare rules ('NEVER do X') and all-caps instead of context and reasoning.
  • Negative-only formatting ('no markdown') instead of describing the target.
  • Putting questions above long documents (20k+ tokens).
  • Assuming prompt engineering alone solves every problem; model choice comes first.

Example

Request explain inflation to my 12 year old

<context>
I want to explain inflation to my 12-year-old, who is curious and reads well but has no economics background. The goal is for them to understand why prices go up over time and why it matters for pocket money and savings.
</context>

<instructions>
Explain inflation in plain language using one everyday example (such as the price of a favourite snack over several years). Then explain one cause of inflation and one effect on savings. Finish with two questions I can ask them to check they understood.
</instructions>

<format>
Write 200-250 words in short, friendly paragraphs suitable to read aloud. Avoid jargon; if you must use a term, define it in the same sentence.
</format>

The Weekly Tally

Useful updates, insights and no-hype advice. Read this week's digest.

One email a week. Unsubscribe in one click.