gemini-prompting

Gemini Prompt Optimization

Guidance for reshaping user requests into effective Gemini prompts. Used by gemini-agent before forwarding to the companion script.

When to reshape

• User request is vague ("look at this code", "help me debug")
• Task would benefit from explicit structure
• Large context needs scoping

When NOT to reshape

• User provides a clear, specific prompt
• Request is simple (e.g., "What is X?")
• User explicitly says "ask Gemini exactly this"

Prompt structure

For complex tasks, use this template:

<task>
[Clear description of what needs to be done]
</task>

<context>
[Relevant files, constraints, prior findings]
</context>

<output_format>
[What shape the answer should take]
</output_format>

Gemini-specific tips

• Gemini has a 1M token context window — use it for large file analysis
• Gemini excels at: code understanding, refactoring suggestions, multi-file analysis
• Be explicit about output format — Gemini tends to be verbose without guidance
• Use "Be concise" or "Reply in under N sentences" for tighter responses
• For code review: specify severity levels and ask for file:line references

Approval modes

Choose the right mode for the task:

• plan — read-only, Gemini can only read files (good for review)
• auto_edit — Gemini can edit files automatically (good for implementation)
• yolo — Gemini auto-approves everything (use sparingly)

Default to plan for analysis, auto_edit for implementation tasks.

Anti-patterns

• Don't send entire codebases as context — scope to relevant files
• Don't ask open-ended questions without constraints
• Don't repeat Claude's own analysis in the prompt (let Gemini give independent opinion)
• Don't include conversation history — Gemini calls are stateless