A multi-role expert panel that reviews a plan before you build.
Uses
APIs to connect
This skill needs keys for these services. The link takes you where to get each one:
What it does
Spawns a panel of sub-agents (architect, qa, judge — always; security / frontend / backend / data / ops — by scope), each on a strict checklist. The judge synthesizes findings into a ranked action list and cross-examines role conflicts. On top — a panel of independent non-Claude judges (OpenAI GPT-5.5 + GLM-5.2 + Kimi K3, optionally DeepSeek / Gemini): their disagreement with Claude is flagged as a blind spot. Roster proven via A/B on 22 plans.
How to use
- 1Say “review this plan” or
/panelwith the plan. - 2The skill picks the roles by scope.
- 3Roles run the plan through checklists.
- 4The judge gives a verdict and priorities (heavy by default).
When it triggers
Example
/panel here's the Supabase migration RFC — assembles architect + qa + data, returns risks and priorities.
Updates
The panel stopped collecting lessons for nothing. Its self-improvement loop had been quietly stalled for two months: it renamed the same weak spot on every run, the repeat counter never added up, and no checklist was ever actually fixed — synonyms are now merged and themes mature again. The health report no longer shows long-closed debt as today's alarm: an alarm you cannot switch off stops being read along with the real ones. And reviews of non-code documents now pull in the legal and unit-economics roles by themselves whenever the text carries obligations between parties, money between parties or personal data — previously they were invited by feel, and on commercial proposals they were systematically missing.
The panel learned to fix itself, not just diagnose. It used to accumulate «this role's checklist has a hole» but was patched by hand — and three themes sealed in July came back, because the clause had been written to the specific examples rather than to the class. Now the scan flags returning themes separately and demands you first explain why the previous wording didn't hold. It also unlocks a self-blocking gate: half the roles had zero entries in the theme dictionary, so every finding got a fresh name and the counter stayed at one forever — the roles that review proposals and estimates had been accumulating findings for nothing. Separately: nobody was checking acceptance thresholds on commercial documents, because the owning role simply doesn't run in that panel. External judges — three non-Claude models — are no longer optional: the step is unconditional, and the skill now notices by itself when they haven't been called in a while.
External judge panel: plans are reviewed by independent non-Claude models (OpenAI GPT-5.5 + GLM-5.2) — where an external judge diverges from the Claude panel, it's flagged as a blind spot. Roster proven via A/B on 22 historical plans: openai+glm give orthogonal coverage, Gemini/DeepSeek turned out redundant (disabled). Independence measured by findings, not verdicts.
A third external judge — Kimi K3 (Moonshot, open-weight 2.8T): added to the panel alongside OpenAI and GLM. Open weights, non-Western lineage → one more independent angle. The integration passed its own finalize review, which caught a real token-budget bug (K3's always-on thinking eats the budget on reasoning before answering) — fixed and live-verified.
Fact-grounder: before the verdict the panel checks the plan's verifiable claims against the FRESH web (via redresearch quick search) and mixes a research_brief into every role and the judge — symmetric to memory. It catches staleness: EOL versions, deprecated APIs, fresh CVEs. Special trigger — when a dependency is installed (npm/pip/cargo…) the latest stable version is verified so nothing outdated gets pinned. Cheap and tier-gated: a plan with no facts makes no external calls.
Added to the RedSkills catalog.
Liked this skill? New breakdowns and updates land in the channel.