Claude prompt best practices for long documents

Claude prefers explicit structure

Anthropic models follow delimiters reliably. Wrap inputs and outputs:

<document>{PASTE_SOURCE}</document>

<task>Summarize risks for an executive audience</task>

<format>
- 5 bullets max
- Each bullet: risk → impact → mitigation
</format>

Sonnet vs Opus

• Sonnet — daily writing, code review, fast iteration
• Opus — multi-step research, adversarial review, complex planning

Use Opus templates only when latency and cost are acceptable; many Sonnet prompts upgrade cleanly by adding "think step-by-step internally, output final only."

Document-grounded tasks

When citing uploaded PDFs or pasted text:

• Tell Claude what to do if the answer is not in the document
• Request quoted spans or paragraph references
• Forbid inventing statistics

Safety and tone

Claude responds well to tone matrices ("confident but not salesy"). Avoid asking for disallowed content; reframe red-team tasks as defensive security reviews with scope limits.

Templates on My Copyprompt

Explore Claude Sonnet and Claude Opus prompts — each includes sample output so you can compare depth before running.

Contribute back

If you maintain Claude prompts for legal review, product specs, or education, submit them publicly with a curator note describing attachment size limits that worked for you.