| name | brains-trust |
| description | Use after any non-trivial build, before commit or deploy. Convene two or more frontier models from different families and providers to review the change for Critical/High/Medium/Low findings; fix cross-validated Criticals before commit and Highs before deploy. Also when the same bug won't fix on the first attempt, or for a whole-codebase pass before a release. |
Brains-trust: a panel of other models before you commit
Your own review pass, and the model that wrote the code, share blind spots. A panel of two or more frontier models from different families and providers catches what a single pass confidently waves through: authz holes, race conditions, platform gotchas. The signal to trust is when reviewers independently flag the same thing. (Even a brand-new flagship benefits, no model escapes its own blind spots.)
Convene the panel. Pick a current frontier model from each major family and run them via different providers, so they're genuinely independent eyes and not the same model twice: for example Anthropic Opus, OpenAI GPT-5.x, Google Gemini 3.x, and strong open models (DeepSeek, Qwen) via OpenRouter, Cloudflare Workers AI, or NVIDIA NIM. Verify the current model ids first (see verify-current).
Brief them. Bundle the changed source plus the goal plus the proof-of-done (or the audit doc) into one prompt. Ask each for findings ranked Critical / High / Medium / Low, each with file-and-line evidence.
Decide by cross-validation, not vote count.
- A finding two or more reviewers independently raise is real. Fix cross-validated Criticals before commit, cross-validated Highs before deploy.
- A finding only one reviewer raises is a maybe; weigh it on its evidence.
- A claim with no supporting evidence is not a finding. Reject it. Panels produce noise; filtering it is part of the job.
Their findings are input to your judgement, not a to-do list. The panel sees the diff, not the why, your constraints, the project's context, what's actually happening. So even a cross-validated finding gets checked before you act on it: verify the advice against the current docs (verify-current) and judge whether it's right for this project, not just correct in the abstract. A confident, agreed-upon suggestion can still be wrong for you. You own the call; the panel only informs it.
The one trap: reviewers can't see the live world. A "this version/API/model doesn't exist" or "that's deprecated" finding is usually a training-cutoff artifact, and two reviewers agreeing on a stale fact is not evidence, they share the same cutoff. Version and existence claims are verify-live-only: route them to verify-current, never let the panel decide them.
Save the panel. Write the findings, what you fixed, and what you rejected (with reasons) to <project>/.jez/audits/<date>-brains-trust-<topic>.md, so the rationale survives the session.
When to run it: after any non-trivial build before commit; a whole-codebase pass before a release; a targeted pass any time the same bug won't fix on the first attempt.
The failure this prevents: shipping the authz hole, the race, or the platform gotcha that a single reviewer (you, or the model that wrote it) confidently missed. The panel's independent eyes are the cheapest place to catch it.
Going further (optional). This skill is the doctrine, runnable with whatever models you have. If you'd rather a built-out tool that convenes the panel for you — provider wiring, cost-aware routing, the convergence-table synthesis — brainstrust is a public sibling that implements it. The skill stands alone; this is just where to look for the machinery.