gemini-prompting

Gemini Prompting Guide

How to compose effective prompts for Gemini CLI in this plugin.

Core principle

Prompt Gemini like an operator — compact, block-structured with XML tags. One clear task per run. State what "done" looks like.

Default prompt recipe

Combine these blocks for most tasks:

<task>What to do. What done looks like.</task>
<structured_output_contract>Format of the expected output.</structured_output_contract>
<default_follow_through_policy>How to handle ambiguity.</default_follow_through_policy>
<verification_loop>Self-check before finishing.</verification_loop>
<grounding_rules>Only use visible context. No speculation.</grounding_rules>

When to add extra blocks

Situation	Add block
Writing/editing code	<action_safety>, <completeness_contract>
Code review	<operating_stance>, <finding_bar>, <calibration_rules>
Research/investigation	<research_mode>, <dig_deeper_nudge>, <citation_rules>
Long-running tasks	<progress_updates>
External tool calls	<tool_persistence_rules>, <missing_context_gating>

Gemini-specific advantages

• Gemini's 1M context window can handle entire codebases — provide broad context.
• Use --include-directories to focus on specific parts of the repo when needed.

Reference files

• references/prompt-blocks.md — all reusable XML blocks
• references/gemini-prompt-recipes.md — end-to-end templates
• references/gemini-prompt-antipatterns.md — failure modes to avoid