Response Length & Output Management

Response length control prevents two failure modes: responses that are too verbose (wasting context tokens and burying important information in prose) and responses that are too terse (omitting critical details or code). GEMINI.md instructions are the primary mechanism — "be concise," "limit explanations to one sentence per change," "show only the changed lines, not the full file." These instructions shape the model's verbosity consistently across a session.

Max output tokens is the hard limit. When set, the model's response is truncated at the specified token count regardless of whether the response is complete. This is a safety net, not a quality control — a truncated response is almost always worse than a complete one. Use maxOutputTokens to prevent runaway responses in automation, not to control verbosity in interactive sessions. For interactive work, GEMINI.md verbosity instructions produce better results because the model adapts its content to fit naturally rather than being cut off mid-sentence.

## Response Format Rules

### Verbosity
- Lead with the answer. Explain after, only if the change is non-obvious.
- Code changes: show the diff, not the full file.
- When modifying 1-3 lines, show only those lines with 2 lines of context.
- When modifying 10+ lines, show the full modified function.
- Never repeat the question back. Never add a summary if the response is under 20 lines.

### Length Limits
- Explanations: 1-2 sentences unless I ask "explain in detail."
- Code reviews: one comment per issue, max 5 issues per file.
- Architecture discussions: use bullet points, not paragraphs.
- When I say "tldr", compress your response to under 3 lines.

1. Define your verbosity preferences Decide how verbose you want responses for different contexts: code changes, explanations, reviews, architecture. Write explicit rules for each context in GEMINI.md.
2. Add escalation triggers Define keywords that increase or decrease verbosity: "explain in detail" expands, "tldr" compresses, "show full file" overrides the diff-only default. These give you real-time control without editing GEMINI.md.
3. Set maxOutputTokens as a safety net only Use maxOutputTokens in automation and CI to prevent runaway costs. For interactive sessions, rely on behavioral instructions — they produce better results than hard truncation.