Gemini Critique

Run Google's Gemini CLI to get an independent critique of a file (spec, code, etc.).

Pre-flight

Verify gemini is available: !which gemini

If gemini is not found, tell the user to install it:

npm install -g @anthropic-ai/gemini-cli  # or see https://github.com/anthropics/gemini-cli

Usage

/critique:gemini $ARGUMENTS

• •$0 — optional file path to critique (relative or absolute)
• •$1 — optional focus area (e.g., "security", "performance", "UX gaps")
• •--model <model> — override the gemini model for this invocation

All arguments are optional. When no file is given, infer the target from conversation context.

Configuration

Config is resolved with the following precedence (first match wins):

1. •CLI flag (--model) — one-off override
2. •Local config (.claude/skill-configs/gemini/config.local.yaml) — personal/local scope, gitignored
3. •Project config (.claude/skill-configs/gemini/config.yaml) — project scope, committed to repo
4. •Default — gemini-3.1-pro-preview

model: gemini-3.1-pro-preview  # model to use with gemini CLI

See config.example.yaml in the critique plugin's gemini skill for reference.

Setup

1. •Check if $ARGUMENTS contains --model <model>. If so, use that model and skip config lookup.
2. •Check for config files (first match wins):
   • •.claude/skill-configs/gemini/config.local.yaml (local scope, gitignored)
   • •.claude/skill-configs/gemini/config.yaml (project scope, committed to repo)
3. •If no config found, use gemini-3.1-pro-preview as the default.
4. •Set ${MODEL} to the resolved model name.

How It Works

1. •Determine what to critique — use one of these strategies, in priority order:
   • •If $0 is provided, use it as the file path
   • •If the user mentions a file by name or path (e.g., @games/foo/bar.js), use that
   • •If changes were made during the current conversation (specs written, code edited, etc.), critique those — use git diff against the state before the conversation's changes to identify the affected files and pass them to gemini
   • •If the user says something like "critique this" or "get a second opinion", look at the most recent substantive work product in the conversation
   • •If nothing can be inferred, ask the user what to critique
2. •Resolve the working directory (project root)
3. •Build a review prompt tailored to the content type — use the templates in ${CLAUDE_SKILL_DIR}/templates/:
   • •spec-review.md — for specification files
   • •code-review.md — for source code files
   • •diff-review.md — for reviewing recent changes
4. •Run gemini non-interactively in sandbox mode
5. •Present the findings to the user

Invocation

cd <project-root> && \
SEATBELT_PROFILE=permissive-closed gemini \
  -s \
  -m ${MODEL} \
  -o text \
  "<review-prompt>"

Flags:

• •-m ${MODEL} — the resolved model (default: gemini-3.1-pro-preview)
• •-s — enable sandbox mode
• •-o text — output as plain text
• •SEATBELT_PROFILE=permissive-closed — restricts file writes and blocks network access

Safety note: The --sandbox flag combined with SEATBELT_PROFILE=permissive-closed restricts writes and blocks network, but is not as granular as Codex's -s read-only. The gemini process can still read all files in the project directory.

Timeout: 300 seconds

After Running

1. •Read the full gemini output (it may be persisted to a temp file if large)
2. •Present findings to the user in a concise summary
3. •If the user wants to act on the feedback, proceed with edits as a separate step

Notes

• •If gemini errors, check gemini --version and verify authentication
• •Unlike the codex backend, gemini does not support reasoning effort levels
• •Must cd to project root before running — gemini has no -C equivalent