gemini-prompting

Gemini Prompting

Use this skill when gemini:gemini-rescue needs to compose a prompt for Gemini.

Gemini works best with structured, phase-driven prompts. Unlike GPT-5.4 which prefers XML block contracts, Gemini responds well to explicit phase decomposition and thinking-level hints.

Core rules:

• Decompose complex tasks into phases: Research -> Strategy -> Execution.
• State the task, expected output shape, and completion criteria upfront.
• Add grounding and verification rules for tasks where hallucination hurts quality.
• Prefer tighter prompt contracts over raising thinking budget.
• Use Markdown sections (not XML tags) as the primary structural unit -- Gemini's training data is Markdown-heavy.

Default prompt recipe:

• ## Task: the concrete job with repository/failure context.
• ## Output Contract: exact shape, ordering, and brevity requirements.
• ## Completion Criteria: what "done" looks like.
• ## Follow-Through Policy: what Gemini should do by default instead of asking.
• ## Verification: required for debugging, implementation, or risky fixes.
• ## Grounding Rules: required for review, research, or claims-heavy tasks.

When to add sections:

• Coding or debugging: add Verification, Completion Criteria, and Missing Context.
• Review: add Grounding Rules, Output Contract, and Dig Deeper.
• Research: add Research Mode and Citation Rules.
• Write-capable tasks (--yolo): add Action Safety.

Model awareness:

• Gemini 2.5 Pro: best for deep analysis, code review, multi-file changes. Supports thinking budget (thinkingBudget: 8192 default).
• Gemini 2.5 Flash: fast iteration, simple fixes, status checks. Use when speed matters more than depth.
• Gemini 3 Pro (preview): strongest reasoning. Use thinkingLevel: HIGH for complex root-cause analysis.
• Do not specify model unless the user explicitly asks. Gemini CLI auto-routes.

Working rules:

• Keep prompts concise. Gemini penalizes prompt bloat more than GPT models.
• Prefer Markdown headers and bullet lists over XML tags.
• Minimize filler and scene-setting (<3 lines).
• One task per prompt. Split unrelated asks into separate sessions.
• For follow-up (--resume), send only the delta instruction.

Prompt assembly checklist:

1. Define the exact task and scope in ## Task.
2. Choose the smallest output contract that makes the answer usable.
3. Decide follow-through: keep going by default, or stop for missing context?
4. Add verification/grounding only where the task needs them.
5. Trim redundant instructions before sending.

Reusable blocks live in references/prompt-blocks.md. Task recipes live in references/gemini-prompt-recipes.md. Common failure modes to avoid live in references/gemini-prompt-antipatterns.md.