Cache Common Prompts

Good prompts are assets. Once you find a prompt that works, save it, version it, and reuse it. Do not retype it from memory every time.

Most agent work is not one-of-a-kind. You ask for the same kinds of things repeatedly:

• Summarize this article
• Review this diff
• Generate unit tests
• Rewrite this for clarity
• Extract structured data

Each of these can become a template with placeholders. Templates reduce variability, improve quality, and lower cost by cutting down on failed attempts.

1. Store them in files. One prompt per file in a prompts/ directory.
2. Use placeholders. Wrap variable parts in {{braces}}.
3. Version them. Name files with a version or date: summarize-v2.txt.
4. Tag them. Add metadata for category, model, and use case.
5. Review them. Update templates when you find a better phrasing.

You are a careful code reviewer. Review the following diff for bugs, security issues, and style violations. If you find a real problem, explain it and suggest a fix. If something is fine, say so briefly. Do not nitpick formatting unless it hurts readability.

DIFF:
{{diff}}

Output your review as a numbered list.

• In your project's repo under prompts/
• In a shared Notion or Confluence page
• In an internal tool or agent platform
• In the agent's own memory if it supports prompt libraries

• Inconsistent outputs from slightly different phrasing
• Wasted tokens on retyping
• Knowledge loss when a team member leaves
• Difficulty debugging why yesterday's prompt worked and today's does not