plan-review

name	plan-review
description	Use when the user wants to review an implementation plan with multiple manually selected AI models before coding. Triggers include multi-model plan review, cross-model consensus, or a judge + verifier pass before implementation handoff.

Plan Review

Use this skill to run a manual multi-model review of an implementation plan.

The skill is model-agnostic. Do not choose models, switch models, call model APIs, or manage API keys. The user manually switches models. The active agent handles only the current step.

Commands

The Python CLI is the reliable command surface:

python3 scripts/plan_review.py start <plan-file.md>
python3 scripts/plan_review.py prepare-review --id <reviewer-name>
python3 scripts/plan_review.py save-review --id <reviewer-name> --content-file <file>
python3 scripts/plan_review.py prepare-judge --id <judge-name>
python3 scripts/plan_review.py save-judge --id <judge-name> --content-file <file>
python3 scripts/plan_review.py prepare-verify --id <verifier-name>
python3 scripts/plan_review.py save-verify --id <verifier-name> --content-file <file>
python3 scripts/plan_review.py status
python3 scripts/plan_review.py list

If slash commands are supported in the host, map /plan-review ... to the same CLI commands.

Flags

• --base-dir <path> — write runs somewhere other than the current working directory. Pass it before the subcommand: python3 scripts/plan_review.py --base-dir /path start plan.md.
• --run <run-id> — operate on a non-current run. Applies to prepare-*, save-*, and status. Omit it to target the run named in plan-review-runs/current.
• --overwrite — replace an existing review, judge, or verifier entry with the same --id. Without it, a duplicate --id fails.
• --content-file <file> — read the answer from a file. Omit it to read from stdin.

Workflow

1. Start a run from a Markdown plan.
2. Ask the user to switch to a reviewer model.
3. Run prepare-review --id <name>.
4. Read the generated prompt path printed by the CLI.
5. Answer the prompt as the active model.
6. Save the answer with save-review --id <name>.
7. Repeat review for each model.
8. Ask the user to switch to the judge model.
9. Run prepare-judge --id <name>, answer it, then save with save-judge --id <name>.
10. Ask the user to switch to the verifier model.
11. Run prepare-verify --id <name>, answer it, then save with save-verify --id <name>.

Run outputs are saved under plan-review-runs/ in the current working directory by default. Prompt templates are loaded from this skill folder.

Rules

1. Always save outputs into the current run folder.
2. Never overwrite review, judge, or verifier files unless --overwrite is passed.
3. Treat --id as a human-readable label only.
4. Use judge/final-revised-plan.md as the source of truth after judging.
5. The final verifier must include one line: Verdict: approved, Verdict: needs_revision, or Verdict: blocked.
6. Do not pass generated run folders into commits unless the user explicitly asks.