Advisor Board

Advisor Board

N persona subagents (N = 4 to 12, configured in references/user-config.md) debate a question across 3 rounds. Clarity over comfort.

Step -1 — First-run check

If references/user-config.md does NOT exist → read and follow references/onboarding.md before anything else. Onboarding captures the user's name, confirms the board roster, records optional context paths, and self-updates {{USER_NAME}} placeholders throughout the skill.

If references/user-config.md exists → read it to get the user's name, board order, and context paths. Proceed.

Step 0 — Context Extraction

Do not assume anything is loaded. Read the question first, then grep the user's context files (from user-config.md) for keywords that match. Pull matching sections only — never full files.

Produce a Context Brief (max 300 words): values/red lines relevant to the decision, current-state items directly involved, strategy elements that apply, behavioral notes only if they explain a pattern in the question, prior attempts if any.

If the user has no context files configured, the Context Brief is built from the question + any conversation history.

Step 1 — Frame the Question

Present this to {{USER_NAME}} and wait for confirmation. This is the only pause point in the main flow.

FRAMED QUESTION:

Situation: [plain-language statement of the decision]

Context that matters:
[Context Brief — max 300 words]

For the board: [crisp, specific version of the question]
What would help most: [stress-test / direction clarity / tradeoff map / action recommendation]

Step 2 — Create Debate File

Path: work/outputs/advisor-board-YYYY-MM-DD-[topic-slug].md (topic-slug = 2-4 word kebab-case summary, e.g. launch-cohort-now, raise-prices, take-xyz-deal)

Sections: ## Framed Question, ## Round 1 — Initial Takes, ## Round 2 — Rebuttals, ## Round 3 — Final Positions & Recommendations, ## Synthesis.

Steps 3-5 — Three rounds of debate

Read references/agent-prompts.md for the three round prompt templates. Let N = the number of board members in user-config.md (between 4 and 12). For each round:

1. Dispatch all N board members (in the canonical order from user-config.md) as subagents in parallel — single message, N Agent tool calls, model sonnet
2. Each agent writes to a temp file: work/outputs/[debate-slug]/r[ROUND]-[name-slug].md
3. Wait for all N confirmations. If any agent returns an ERROR (e.g. persona not found), stop and report to {{USER_NAME}} before proceeding
4. Orchestrator assembles temp files into the debate file in canonical order, then deletes the temp dir
5. Only after assembly is complete, proceed to the next round

Rounds are sequential across, parallel within — Round 2 starts only after all Round 1 temp files are assembled.

No substitutions, no invented members. If an agent cannot find their ## [NAME] section in references/board-members.md, it must stop and return an ERROR. Never fabricate a persona.

Step 6 — Synthesis

Read the complete debate file. Write the synthesis directly into the file:

## Synthesis

### Where the board converges
[2-4 bullets — only list genuine alignment across a majority of members. Do not pad.]

### Unresolved tensions
[2-4 real disagreements that stayed disagreed after Round 3. Name specific members on each side. State the core disagreement plainly.]

### Clearest paths forward

**Path A: [name]** — what this path is, which members point toward it, what it requires accepting or giving up.

**Path B: [name]** — same format, only if genuinely distinct from Path A.

**Path C: [name]** — same format, only if applicable.

### The real question underneath this
[One sentence. The decision the debate revealed underneath the surface question.]

Step 7 — Close

Report the debate file path. Offer to go deeper on any specific member's reasoning.

Rules

• All subagents use model sonnet
• Board roster is fixed per user-config.md (N between 4 and 12) — no substitutions, no additions, no invented members during a debate
• Each agent writes its own temp file per round; orchestrator assembles after all confirmations — no concurrent writes to the debate file
• Orchestrator re-reads files fresh each session — never assume context is already loaded between rounds
• Context Brief is max 300 words, built from grep hits — never pass raw full files to agents
• Each agent greps for its own ## [NAME] section in board-members.md and reads only that section (offset/limit), never the full file
• First person only — agents speak AS the persona, never "Goggins would say"
• No softening, no manufactured consensus — preserve real disagreement through to the synthesis
• Parallel within rounds, sequential across rounds