council-review

name

council-review

description

Pressure-test any project, system, document, or plan from multiple independent expert perspectives at once, then synthesize one prioritized report. Convenes a "council" of reviewer agents — each assigned a distinct lens (e.g. brand/voice, content/editorial, technical reliability, distribution/growth, operations/ maintainability) — that review in parallel, then a chair synthesizes their findings, surfacing agreements and conflicts. Use when the user asks for a "council review," "multi-perspective review," "pressure test," "review this from different angles," "panel review," "get multiple perspectives on," "red-team this," or wants a project analyzed from several points of view. Report-only by default — it recommends, it does not change anything.

Council Review

Convene a council of independent expert reviewers to pressure-test a subject from several distinct points of view, then synthesize their findings into one prioritized report. The value is independent perspectives that don't contaminate each other plus a synthesis that surfaces where they agree (high-confidence) and where they conflict (the interesting tensions).

This skill is report-only. It recommends; it does not modify the subject. If the user wants changes applied afterward, that's a separate, explicit follow-up.

When to use

• "Run a council review on X" / "pressure-test this" / "review this from different angles"
• Evaluating a routine, codebase, feature, document, plan, strategy, PR, or design
• Any time multiple independent lenses would catch more than one reviewer would

Mechanism

This skill drives the Workflow tool to fan out one agent per lens (parallel, independent), each returning structured findings, then the chair (you) synthesizes. Invoking this skill is explicit opt-in to multi-agent orchestration — go ahead and call Workflow. If Workflow is unavailable, fall back to reviewing each lens yourself in sequence and synthesizing, but say so.

Step 1 — Scope the subject

Identify exactly what is under review and gather the inputs the reviewers need:

• A path / directory / file set / PR / pasted artifact. Resolve it concretely.
• Read enough yourself to know the subject's type (content routine? service codebase? strategy doc? product spec?) — this drives lens selection in Step 2.
• Decide how reviewers will access the subject:
  • File-readable (default): pass each agent the key paths and let it read them.
  • Not file-readable (e.g. an isolated subprocess, a pasted blob, a remote artifact): embed the full subject text into each agent's prompt. Reviewers can't review what they can't see.

Step 2 — Choose the lenses

Pick 3–6 lenses tailored to the subject. Each lens must be genuinely distinct — overlapping lenses waste agents. The user may name lenses or a focus; honor that. Otherwise select from a pack below or compose your own.

Canonical pack — content / marketing routine (the proven default mix):

1. Brand, positioning & voice — does it reinforce the intended brand and audience?
2. Content sourcing & editorial quality — is the input selection and output quality good and sustainable?
3. Technical reliability & failure modes — where does it break, especially unattended?
4. Distribution & growth strategy — are the channels, formats, timing, and reach choices right?
5. Operations & maintainability — is it robust, observable, and cheap to maintain over time?

Pack — software / codebase: correctness & bugs · security · performance & scalability · architecture & maintainability · testing & coverage · developer experience / API design.

Pack — product / strategy / business plan: market & positioning · customer & demand evidence · business model & unit economics · competitive & differentiation · execution & operational feasibility · risk & failure modes.

Pack — document / writing / proposal: argument & logic · evidence & sourcing · clarity & structure · audience fit & voice · completeness & gaps · counterarguments / red-team.

Adapt freely: drop lenses that don't apply, add subject-specific ones (regulatory, legal, accessibility, cost, etc.). Scale lens count to the user's ask — a quick pressure test is 3 lenses; "be thorough" is 5–6.

Step 3 — Run the council (Workflow)

Fan out one agent per lens in parallel, each with the shared structured schema below, then return the raw findings. Template — adapt the lens list, subject context, and how reviewers access the subject:

export const meta = {
  name: 'council-review',
  description: 'Multi-perspective council review of <subject>',
  phases: [{ title: 'Council review' }],
}

const SUBJECT = `<one-paragraph description of what's under review + how to access it:
either "read these paths: ..." or the embedded artifact text itself>`

const SCHEMA = {
  type: 'object', additionalProperties: false,
  required: ['lens', 'overall_take', 'strengths', 'issues', 'top_recommendations'],
  properties: {
    lens: { type: 'string' },
    overall_take: { type: 'string', description: '2-4 sentence verdict from this lens' },
    strengths: { type: 'array', items: { type: 'string' } },
    issues: { type: 'array', items: {
      type: 'object', additionalProperties: false,
      required: ['severity', 'title', 'detail', 'recommendation'],
      properties: {
        severity: { type: 'string', enum: ['high','medium','low'] },
        title: { type: 'string' },
        detail: { type: 'string' },
        recommendation: { type: 'string' },
      } } },
    top_recommendations: { type: 'array', items: { type: 'string' },
      description: 'ranked: the 2-4 changes this lens would prioritize' },
  },
}

const base = `You are one member of an expert review council evaluating the following subject.
This is a REVIEW ONLY — recommend, do not change anything. Be specific: cite the exact
mechanism/line/section, distinguish real problems from nitpicks, and make recommendations
actionable. Disagreement with current design choices is welcome if justified.

SUBJECT:
${SUBJECT}`

const LENSES = [
  { label: 'review:lens-1', prompt: `${base}\n\nYOUR LENS — <name>: <2-4 sentences of exactly what to evaluate from this angle, with pointed questions>` },
  // ... one entry per lens
]

const findings = await parallel(LENSES.map(l => () =>
  agent(l.prompt, { label: l.label, phase: 'Council review', schema: SCHEMA })
))
return findings.filter(Boolean)

Each lens prompt should: (a) carry the shared base context, (b) state the lens name, and (c) give 2–4 sentences of pointed, lens-specific things to evaluate plus guiding questions. Sharper lens prompts → sharper findings.

Step 4 — Synthesize (the chair)

You are the chair. Read all returned findings and write ONE report. Do not just concatenate the agents:

• Cross-cutting themes — issues ≥2 lenses raised independently are high-confidence; lead with these.
• Conflicts — where lenses disagree, surface the tension and take a position (more recent / better-evidenced wins); do NOT average them away.
• Prioritized recommendations — a single ranked list, highest-leverage first, each with what-to-change and why. Pull severity from the agents but re-rank holistically.
• Keep what's working — name the strengths worth preserving so recommendations don't read as "rewrite everything."

Output format

# Council Review — <subject>

## Verdict
<2-4 sentences: is this well-designed? overall health.>

## Council
<one line per lens: name + its headline take>

## Strengths to preserve
- ...

## Cross-cutting findings (multiple lenses agreed)
- ...

## Conflicts & tensions
- <where lenses disagreed and your call>

## Prioritized recommendations
1. **<change>** (severity, lens) — what to change and why it matters.
2. ...

## Per-lens detail
<collapsible/short per-lens summaries for the reader who wants depth>

Notes

• Report-only. Never edit the subject as part of this skill. Offer a follow-up if the user wants fixes applied.
• Scale to the ask. Quick pressure test = 3 lenses; "thorough"/"red-team" = 5–6 plus sharper, more adversarial lens prompts.
• Independence is the point. Don't let one lens's prompt reference another's expected findings; they review blind and the chair reconciles.
• Token awareness. Each lens is a full agent; 5–6 lenses on a large subject is a meaningful spend. Match lens count to how much the decision warrants.