redplan

Multi-role plan verification + code-review skills for Claude Code. A panel of expert subagents (architect, qa, security, backend, ops, data, frontend, judge) reviews your plan before you build — or your diff before you merge — with strict checklist-driven protocols and a synthesizing judge. Optional cross-model second opinion via GPT-5 + Gemini 2.5 Pro.

This repo ships two sibling skills that share one role engine:

Skill	Input	Use	Verdict
plan-panel (/plan-review)	a plan / RFC / task	catch design mistakes before coding	priority-ranked action list
finalize (/finalize)	the working git diff	stabilize (typecheck/lint/test) + review before merge	SHIP / FIX-FIRST / NEEDS-WORK

---

What problem does it solve

The plan is where ~80% of mistakes are baked in (missing acceptance criteria, security vectors, broken assumptions, premature abstractions), and the diff is where the rest sneak in. One LLM reading its own work rarely catches these — there's no adversarial pressure.

redplan makes it adversarial: multiple expert roles review in parallel with structured protocols (checklist + severity rubric + JSON schema), a judge synthesizes with conflict resolution, and optionally a second-opinion pass through other models finds Claude-specific blind spots. finalize reuses the same roles in review_mode=code against a real diff.

---

Quick start

Install (1 command — installs both skills)

curl -fsSL https://raw.githubusercontent.com/Redloft/redplan/main/install.sh | bash

The installer checks deps (git, curl, jq, node ≥18, python3), copies plan-panel/ → ~/.claude/skills/plan-panel, finalize/ → ~/.claude/skills/finalize (sharing plan-panel/lib via relative symlinks), installs the slash-commands, and runs doctor.sh.

Manual install:

git clone https://github.com/Redloft/redplan /tmp/redplan
cp -R /tmp/redplan/plan-panel  ~/.claude/skills/plan-panel
cp -R /tmp/redplan/finalize    ~/.claude/skills/finalize
cp /tmp/redplan/plan-panel/commands/*.md ~/.claude/commands/

Or paste in any Claude Code session — Claude reads this README and walks you through:

Install this skill: https://github.com/Redloft/redplan

Use

/plan-review <your plan as text>      # verify a plan
/plan-review --from-task <task>       # draft a plan, then loop-review it (experimental)
/finalize                             # stabilize + review the current working diff
/finalize --since <ref>               # review a diff against a ref

Diagnose / update / uninstall

bash ~/.claude/skills/plan-panel/lib/doctor.sh        # health-check (deps, optional --ultra setup)
curl -fsSL .../install.sh | bash                      # re-run = update (idempotent)
rm -rf ~/.claude/skills/plan-panel ~/.claude/skills/finalize ~/.claude/commands/{plan-review,panel-*}.md

---

How it works

1. scope — a scoper reads the plan/diff and picks which roles apply.
2. roles — each selected role reviews in parallel under a strict protocol (checklist → findings with severity → JSON schema). plan-panel reviews a plan; finalize reviews a diff (review_mode=code) after a deterministic stabilize pass (typecheck/lint/test + bounded auto-fix).
3. judge — synthesizes all findings, resolves conflicts, emits a priority-ranked action list and a verdict.
4. (optional) cross-model — for --ultra, re-run critical findings through GPT-5 / Gemini 2.5 Pro and meta-judge the disagreements.

Roles, the workflow orchestrators (workflow/*.js), and the shell lib are all in this repo. Self-improvement: /panel-feedback + /panel-solidify refine role prompts from accumulated feedback (human-in-loop, prompts only).

Optional: cross-model verify (--ultra)

Needs OPENAI_API_KEY / GEMINI_API_KEY in the environment, or 1Password CLI refs (configurable). Without them, redplan runs Claude-only and says so. See plan-panel/lib/cross-model.sh and doctor.sh.

License

MIT © 2026 Igor Konovalchik