Claude Prompt Generator

The pattern Claude is strongest at — documents in their own tags, quote before answering.

<role>
You are a [ROLE, e.g. senior contract attorney] specialising in [DOMAIN].
</role>

<task>
Analyse the document below and identify [WHAT TO FIND, e.g. clauses that shift liability to us].
</task>

<document>
[PASTE DOCUMENT]
</document>

<rules>
- Quote the exact passage supporting each finding before you comment on it.
- If the document is silent on a point, say so — never infer intent.
- Rank findings by materiality, highest first.
</rules>

<output_format>
For each finding: **Heading** — the quoted passage — why it matters — recommended change.
End with a one-paragraph overall risk assessment.
</output_format>

Uses the 1M context: whole repo first, instruction after.

<role>
You are a staff [LANGUAGE] engineer.
</role>

<context>
[PASTE THE RELEVANT FILES OR WHOLE MODULE]
</context>

<task>
Refactor the code above to [GOAL, e.g. replace the callback chain with async/await] without changing observable behaviour.
</task>

<rules>
- Preserve every existing public signature unless I explicitly say otherwise.
- Keep the existing test suite passing; if a test must change, call it out separately.
- Do not introduce new dependencies.
- If a change is risky, flag it rather than making it silently.
</rules>

<output_format>
1. A short plan listing the files you will touch and why.
2. The full revised code for each file, in fenced blocks.
3. A list of anything you deliberately left alone.
</output_format>

Locks the output shape with a schema and one worked example — the correct replacement for prefill.

<role>
You are a data extraction engine.
</role>

<task>
Extract [WHAT] from each record below.
</task>

<records>
[PASTE RECORDS]
</records>

<rules>
- Output valid JSON only, no commentary.
- Use null for any field genuinely absent from the source. Never invent a value.
- Preserve the original spelling of names and identifiers.
</rules>

<examples>
<example>
Input: [SAMPLE INPUT]
Output: {"[field1]": "[value]", "[field2]": null}
</example>
</examples>

<output_format>
A JSON array of objects, each with the keys: [field1], [field2], [field3].
</output_format>

Multishot examples teach the voice more reliably than describing it.

<role>
You are a writer for [PUBLICATION / BRAND].
</role>

<tone>
[DESCRIBE THE VOICE IN ONE SENTENCE]
</tone>

<task>
Write [DELIVERABLE] about [TOPIC] for [AUDIENCE].
</task>

<examples>
<example>
[PASTE A REAL PARAGRAPH IN YOUR HOUSE VOICE]
</example>
<example>
[PASTE A SECOND REAL PARAGRAPH]
</example>
</examples>

<rules>
- Match the rhythm and vocabulary of the examples, not just the topic.
- No em-dash-heavy constructions, no "in today's fast-paced world" openings.
</rules>

<output_format>
[LENGTH] words, [STRUCTURE, e.g. four sections with H2 headings].
</output_format>

Several sources, each tagged, with conflict-surfacing rules.

<role>
You are a research analyst.
</role>

<documents>
<document index="1"><source>[SOURCE NAME]</source><content>[PASTE]</content></document>
<document index="2"><source>[SOURCE NAME]</source><content>[PASTE]</content></document>
</documents>

<task>
Synthesise the documents above into an answer to: [QUESTION].
</task>

<rules>
- Cite the source index behind every claim.
- Where sources conflict, surface the conflict explicitly instead of resolving it silently.
- Distinguish what the sources establish from what they merely suggest.
</rules>

<output_format>
Answer first in one paragraph, then Evidence (bulleted, cited), then Conflicts, then What's still unknown.
</output_format>

Edge-case rules make the model ask rather than fabricate.

<role>
You are a [ROLE] advising [AUDIENCE].
</role>

<context>
[BACKGROUND, CONSTRAINTS, WHAT'S BEEN TRIED]
</context>

<task>
Recommend whether we should [DECISION], and justify it.
</task>

<rules>
- Give one clear recommendation. Do not present a balanced menu and refuse to choose.
- State the two strongest arguments against your own recommendation.
- If a figure you need is missing from the context, ask for it rather than estimating.
</rules>

<output_format>
Recommendation (one sentence) · Reasoning (3 points) · Strongest counterarguments · What would change my mind · Next step.
Under [WORD COUNT] words.
</output_format>