# Counsel > Convene a two-vendor counsel — flagship reviewers from different vendors at high effort — to adversarially review an important architecture/design decision or proposal before committing. Use when the user explicitly asks for counsel or a two-vendor adversarial review. - Skill: `jorgemenadev/counsel` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add jorgemenadev/counsel` - Raw SKILL.md: https://api.skillmd.com/api/skills/jorgemenadev/counsel/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: jorgemenadev (https://skillmd.com/u/jorgemenadev) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/jorgemenadev/counsel --- # Counsel (portable protocol) Two flagship reviewers from **different vendors** attack a written proposal in parallel; you **chair**: synthesize their findings, revise, re-convene until verdicts converge — **max 3 rounds**. The chair rules on every finding and owns the final call; the seats advise. Runtime launch mechanics live in adapters — not here: - Claude Code: consumer-local `.claude/skills/counsel/SKILL.md` - Codex / GPT chair notes: `references/codex-chair.md` - GitHub Copilot CLI: `references/github-copilot-chair.md` ## Naming (do not overclaim) | Label | When | |---|---| | **counsel** | Two vendors successfully participated | | **second opinion** | Only one vendor/family participated (engine down or unavailable) | | **OpenAI counsel-style review** | Multiple OpenAI reviewers, no Anthropic seat | ## Contract - Two seats, two vendors, every round. An engine down ⇒ degrade loudly: a one-seat run is a *second opinion*, not a counsel — name that in the record. Each seat's vendor must be established by evidence the chair can verify (runtime metadata, launch surface), **not** by the seat self-reporting its own model — where one tool launches both seats and only a model string differs, self-report can silently echo a substituted default and fake the second vendor. - The proposal is a **file**, and it names the source paths reviewers verify claims against — reviewers judge evidence, not prose. - Every proposal declares its **altitude** — ordered `idea → plan → spec → implementation` — and lists what is deliberately deferred below it. Seats judge at that altitude: a listed deferral is judged for *safety to defer*, never as an absence, and a below-altitude finding is appendix material, not a verdict driver. - Every finding gets a **disposition**: accepted → visible revision in the next draft; rejected → one-line justification in the record. A finding about a *fact* (proposal contradicts a file) is settled by the chair re-reading the file, never by preference. - Seats read; only the chair writes the proposal. - Secrets never enter prompts or reports — existence checks only. ## Bench (logical seats) Edit this table when flagships change. Adapters map these seats onto concrete launch tools. | Seat | Vendor family | Role | |---|---|---| | openai | OpenAI flagship (default `gpt-5.6-sol`, high effort) | Adversarial reviewer | | anthropic | Anthropic flagship (default `fable-5`, high effort) | Adversarial reviewer | ## The loop 1. **Setup** — create a session scratchpad dir; state comes from disk, not memory: ```bash DIR="/counsel/" && mkdir -p "$DIR" ROUND=$(( $(ls "$DIR"/proposal-v*.md 2>/dev/null | wc -l | tr -d ' ') + 1 )) echo "DIR: $DIR"; echo "ROUND: $ROUND" ``` 2. **Draft** `proposal-v$ROUND.md`: first `Altitude: — deferred: `, then the decision, verified current state, design, alternatives, risks, sizing, and source paths. Round ≥2: append prior findings + disposition table and "what changed in this draft". 3. **Convene both seats in parallel**, same reviewer prompt (below), each blind to the other. Round 1 spawns fresh; rounds ≥2 resume each seat's own session when the runtime supports it (else relaunch fresh — appendix protocol makes both modes equivalent). 4. **STOP — wait for both reports before writing a word of synthesis.** Ruling on the early return anchors the round on one vendor and reduces the counsel to an echo. 5. **Synthesize**: read both `## Plan verdict` sections first; they frame the round. A valid report has that as its first substantive section (runtime attestation metadata may precede it), answers all five questions, has ≤7 decision-ranked findings, gives below-altitude notes no severity, and traces its final verdict to one plan-verdict answer. Return a malformed report to its seat once; a second malformed result is an invalid seat, so apply the naming table before synthesis. Then merge findings, mark agreements/conflicts, and write dispositions only for findings at or above altitude. Verdict gate: - Both `SHIP AS-IS`, or every accepted change already incorporated → **CONVERGED**, go to 6. - Any accepted BLOCKER or a valid `RETHINK` → material revision: back to 2. A `RETHINK` driven only by below-altitude notes is malformed. - `ROUND` = 3 → **EXHAUSTED**: chair rules; standing objections recorded as **DISSENT**. - Findings all rejected with justification → do not re-convene; rule now. 6. **Deliver** the counsel record (format below) ending with "Your actions". Decision that sticks → vault ADR/entity note per vault rules, linking the record. ## Reviewer prompt (both seats — fill `{}`) > You are one seat of a two-vendor counsel reviewing a proposal whose header declares its **altitude** and deliberate deferrals. Be adversarial and concrete — verify every claim you can against the sources; do not praise. > Read `{proposal path}`. Verify against: `{source paths}`. Never read or quote credential values — existence checks only. > {Round ≥2: Prior findings and dispositions are in the appendix; re-litigate a rejection only with new evidence. Answer the plan verdict fresh — dispositions feed findings, never replace whole-proposal judgment.} > Report, in this order: > 1. `## Plan verdict` — prose answering: Is this the right problem? Is the approach sound and proportionate? What is the single most important thing? What simpler version would you build and what would it lose (`none` is valid with reasons)? Which listed deferrals are unsafe? This must be the first substantive section. > 2. `## Findings` — max 7, ranked by decision impact, each `severity (BLOCKER/MAJOR/MINOR) — claim it concerns — concrete fix`; only findings at or above the declared altitude belong here. > 3. `## Below altitude` (optional) — unranked, severity-free notes for later stages; they never justify the verdict. > End with one verdict line — `SHIP AS-IS` / `SHIP WITH CHANGES:` (list) / `RETHINK:` (why) — naming which plan-verdict answer drives it. > Write the report to `{DIR}/review--r{ROUND}.md` when the runtime can write files; otherwise return the report as the final message for the chair to save. ## Counsel record (deliverable to Jorge) ``` ## Counsel: — after N round(s) | Round | openai verdict | anthropic verdict | Material changes | Findings that changed the design: Rejected findings + justification: Dissent (if any): Final proposal: ``` If degraded: title the record `## Second opinion: …` (or `## OpenAI counsel-style review: …`) and say which seat was unavailable. ## Anti-patterns - **Anchoring** — synthesizing from the first report while the second runs. - **Rubber-stamp round** — re-convening with no material revision. - **Chair abdication** — pasting a seat's review as the decision. - **Echo bench** — both seats from one vendor because the other engine was down, still calling it counsel. - **Altitude creep** — auditing schema, parser, payload, or lock minutiae below the declared altitude. Those notes belong below altitude; dispositions for them drag the next round down a level. - **Shrinking agenda** — replacing fresh whole-proposal judgment with prior detail dispositions. Every round re-answers the plan verdict against the current draft.