Council

Council

Analyzes a task from multiple expert perspectives independently, then synthesizes a single reconciled plan.

/council {task}                              — Auto-select relevant personas (4-7)
/council {task} --personas arch,pragma,dx    — Explicit subset
/council {task} --subroutine                 — Non-interactive (for /plan and /q callers)

---

Personas

| Key | Persona | Lens | Standard | |-----|---------|------|----------| | arch | Fundamentals Architect | Type safety, abstractions, API contracts | "Would this survive a major refactor?" | | pragma | Pragmatic Maximizer | Ship value, scope control, reversibility | "Smallest change for 80% of the value?" | | dx | DX Perfectionist | API ergonomics, discoverability, errors | "Would a new dev understand without reading source?" | | ux | UX Advocate | A11y, interaction quality, perceived perf | "Does this feel right for every user, every device?" | | design | Design Purist | Design system coherence, tokens, rhythm | "Does this respect the system's visual language?" | | perf | Performance Engineer | Bundle size, runtime cost, rendering | "What's the cost at 1000 components, slow 3G?" | | maint | Maintainability Guardian | Readability, testability, coupling | "Will the next person curse or thank us?" |

Conflict weights: pragma on scope, arch on API contracts, perf on measurable perf, ux+design on user-facing, maint on long-term cost, dx on ergonomics.

---

Auto-Selection (when --personas omitted)

• Always: pragma + maint
• If UI/components: + ux, design
• If APIs/types: + arch, dx
• If rendering/bundle: + perf
• If tooling/workflow: + dx
• Default all 7 if broad/ambiguous
• Minimum 4 — fill by: arch → dx → perf → ux → design

---

--subroutine Flag

When called by /plan or /q: skip EnterPlanMode, resolve all open tensions via domain authority + confidence weighting, return plan directly. No AskUserQuestion. The caller handles user interaction.

When NOT set: enter plan mode, present open tensions for user resolution.

---

Phase 1: EXPLORE

Launch one Explore agent (subagent_type: "Explore", thoroughness: "very thorough"):

Explore the codebase for this task: {task}
Produce: 1) Affected files/modules 2) Existing patterns 3) Constraints
4) Prior art 5) Open questions. Check tests, types, and usage sites.

Parallel recall (fire together):

• recall-prior-art with task keywords (project-scoped)
• recall-lateral with task keywords (unscoped — cross-domain serendipity, limit 5)

Inject recall results into context brief as ## Prior Knowledge and ## Lateral Connections sections.

Output: structured context brief fed to all personas.

---

Phase 2: DELIBERATE

Launch all persona agents in parallel (single batch). subagent_type: "Plan", mode: "plan".

Each persona gets identical context + task. Prompt:

You are the {PERSONA_NAME}. LENS: {LENS}. STANDARD: "{QUESTION}"

CONTEXT: {explore_brief}
TASK: {task}

Respond with EXACTLY:
## Proposal — 2-5 bullets, concrete and actionable
## Risks — 1-3 bullets on YOUR OWN proposal
## Non-Negotiables — 1-2 things you won't compromise on
## Concessions — 1-3 things you'd give up if pushed back
## Confidence — 1-5 rating + one sentence why

All personas launch simultaneously. They never see each other's output.

---

Phase 3: RECONCILE

Synthesize directly from persona outputs (no agent). Four filters in order:

3a. Agreements: 3+ personas align → adopt as plan foundation.

3b. Cherry-picks: Unique idea from one persona, unchallenged → adopt with attribution.

3c. Compromises: Resolve conflicts via:

1. Pre-declared concessions
2. Confidence weighting (higher wins)
3. Domain authority (conflict weights above)
4. Synthesis (satisfy both constraints)

Document: "Conflict {A} vs {B} on {topic}: resolved by {reasoning}"

3d. Open Tensions:

• Standard: present both sides, recommend default, mark for user decision
• --subroutine: resolve all via domain authority, document reasoning

---

Phase 4: PRESENT

# Council Plan: {task_summary}

**Personas:** {list with keys}

## Plan
{Numbered actionable steps}

## Key Decisions
{Agreements, cherry-picks, compromises — briefly attributed}

## Open Tensions
{Standard: both sides + recommended default}
{--subroutine: omitted — all resolved under Key Decisions}

## Risk Summary
{Top 3-5 risks, deduplicated across personas}

Standard mode: If tensions → ask user to resolve. Suggest /plan or /q handoff. Subroutine mode: Plan is final. Return to caller.

---

Example

/council "Add dark mode toggle to the app"

→ Auto-selects all 7 (UI task touches all perspectives). Full deliberation. Presents reconciled plan with any open tensions for user resolution.

---

Memory Integration

Recall enriches exploration. Store captures decisions. Read ~/.claude/skills/memory/MEMORY_OPS.md for tool call templates.

| Phase | Action | |-------|--------| | EXPLORE | recall-prior-art + recall-lateral (parallel). Inject as ## Prior Knowledge + ## Lateral Connections. | | DELIBERATE | Per-persona discipline recall: arch → "type safety abstractions", ux → "accessibility interaction", design → "design tokens visual". Others → recall-prior-art with task keywords. | | PRESENT | store-decision with key decisions + rationale | | User resolves tension | store-feedback with the resolution reasoning |

---

Rules

1. Plan mode enforced (standard). EnterPlanMode first. Exception: --subroutine.
2. Personas are independent. Never see each other's output.
3. Reconciliation = ONE plan. Not a comparison table.
4. Concessions make reconciliation tractable.
5. Confidence is honest. 2/5 = "defer to others."
6. Minimum 4 personas. Fewer → just plan directly.
7. pragma + maint always present (unless --personas overrides).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: vlting
• Source: vlting/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.