Advisory Board
advisory-board is a reusable workflow for getting an idea, plan, or decision reviewed by several high-reasoning models before you commit to it.
Bring the board any material worth more than one expert opinion — a plan, design, architecture, document, decision, or goal. Each model reviews it independently, then reads a packet of the first-round findings and responds to the strongest objections, missed details, and points of disagreement. A final synthesis turns the debate into a single working handoff.
See a sample: a rendered handoff from a payments idempotency review — self-contained HTML that opens offline.
What It Is For
- Adversarial review of a plan, design, or architecture before a build.
- Pressure-testing a decision, strategy, or proposal.
- Implementation sequencing.
- Surfacing risks, stale assumptions, and missing evidence.
- Letting several models debate and sharpen each other’s thinking.
- Turning multiple model opinions into a single takeaway.
Default Workflow
- Round 1 — each model reviews independently, with no view of the others.
- Round 2 — each model reads the other seats’ findings and responds: what it missed, what changed its mind, what it still disputes.
- Final synthesis — the chair writes a handoff with consensus, dissent, risks, guardrails, and next actions.
Two rounds is the default: enough for the board to challenge itself without turning every review into a long-running process. Set rounds to auto to stop early on convergence or add a round when real disagreement remains.
Seats
Each seat gets a distinct lens, matched to the subject — ready-made lens sets for software, product, research, legal, business, and writing live in lens-presets.md. For software and technical work:
- Claude — architecture, systems, and adversarial design review.
- Codex — repo-grounded implementation, migration, testing, and execution.
- Gemini — product, operations, rollout, latency, evaluation, and user-workflow risk.
For non-software subjects the seats take comparable lenses — first-principles soundness, execution and feasibility, and second-order consequences. Every seat still reviews the whole brief; the lens reduces blind spots, it doesn’t narrow responsibility.
Safety Defaults
- A preflight go/no-go check (CLI, auth, model, smoke ping) gates every run — a board needs at least two healthy seats.
- Sensitive material is disclosed before it leaves the machine, with redaction or a local-only board as options.
- Read-only unless edits are explicitly requested.
- Subscription-backed CLIs where available.
- Highest available reasoning settings, verified at run time.
- No secrets stored in artifacts.
Outputs
A run saves:
- per-seat Round 1 notes;
- per-seat Round 2 rebuttals;
- a board packet between rounds;
- a final consensus handoff (Markdown + self-contained HTML);
verdict.json — a machine-readable verdict you can gate CI on or reshape into a PR comment, Slack message, or TL;DR;- run metadata: provenance, per-seat status, and timings (no secrets).
Source Files