Council

DecisionCouncil

A disciplined, adversarial decision procedure inspired by Karpathy's LLM Council and the ECC council pattern. One orchestrator (you, the main session) frames the question, spawns sub-agents with distinct lenses, uses last30days and deep-research only where they earn their tokens, seats a Fact Checker, surfaces real disagreement, and synthesizes a ranked list with a clearly-marked #1 plus the condition that would change the answer.

Different from opinion-only councils: Scout establishes a shared baseline once. Fact Checker verifies contested claims (VERIFIED / PARTLY / FALSE / UNVERIFIABLE). Output is a ranked shortlist, not a single blended paragraph.

Companion skills (install together): last30days for recent social/web signal; deep-research for claim verification.

Install: npx skills add Endokelp/DecisionCouncil --skill council -g -y

Cursor (primary)

Install: ~/.cursor/skills/council/SKILL.md (run install.ps1 or install.sh from this repo).

Invoke: user must name the skill or use a trigger phrase (this skill does not auto-load):

• council this: <question>
• /council <question> or /council full <question>
• run the council, pressure-test this, stress-test this, war room this, debate this
• @council or "use the council skill"

Sub-agents: use the Task tool. Spawn each wave in one message with multiple parallel Task calls.

Role	subagent_type	run_in_background	Notes
Scout, Advocate, Skeptic, Fact Checker	generalPurpose	false	Council waves must block until each agent returns
Repo/file discovery only	explore	false	Use when Step 0 needs codebase context, not for debate personas

Task prompt template. Give each sub-agent:

1. Step-0 frame (decision, options, constraints, deciding metric)
2. Prior wave output only (not full chat history)
3. Role instructions + word cap
4. "Return ranked top-N + list of factual claims for Fact Checker"

Example (Advocate + Skeptic in parallel):

Task(description="Council Advocate", subagent_type="generalPurpose", run_in_background=false, prompt="...")
Task(description="Council Skeptic", subagent_type="generalPurpose", run_in_background=false, prompt="...")

Do not spawn subagents for orchestrator synthesis. That stays inline in the main session.

Cursor anti-patterns (never):

• run_in_background: true for council waves. Synthesis depends on completed persona output.
• explore for Advocate/Skeptic/Fact Checker. Debate personas need generalPurpose + web search.
• Multiple sequential messages per wave. Batch parallel Task calls in one orchestrator turn.
• Passing full chat history into sub-agent prompts. Step-0 frame + prior wave only.

Wave checklist (Cursor):

Lite council:
- [ ] Step 0 frame inline (or 1× explore Task if repo context needed)
- [ ] 1× Task Scout
- [ ] 2× Task Advocate + Skeptic (same message)
- [ ] 1× Task Fact Checker
- [ ] Chairman synthesis inline (no Task)

Install via repo: npx skills add Endokelp/DecisionCouncil --skill council -g -y or install.ps1 / install.sh.

Claude Code

Works with /council. Use the Agent tool instead of Task; same wave structure and token caps.

When to use / when NOT to

• USE for: expensive purchase, architecture or vendor lock-in, strategy fork, "which of these options". Anything costly to get wrong or hard to reverse.
• DO NOT use for: reversible, cheap, or low-stakes choices. The council is deliberately heavyweight.

The research model

Each lens argues from its own evidence, not vibes:

1. Scout (1 Task): run last30days when recency/social signal matters, plus lean baseline (option list, prices, headline specs, deciding metric). ≤ ~600 words, URL per number. No opinion.
2. Advocate + Skeptic (2 parallel Tasks): each gets frame + Scout baseline; ≤5 lens-specific web searches for NEW evidence only.
3. Fact Checker (1 Task): run deep-research in Quick mode on contested claims. VERIFIED / PARTLY / FALSE / UNVERIFIABLE + cited source. ≤ ~700 words.
4. Orchestrator synthesizes inline as Neutral + Pragmatist + chairman. No extra agent.

Do not invoke last30days or deep-research inside Advocate/Skeptic. Those personas stay search-light.

Token discipline (mandatory)

1. Baseline once, angles many. Scout establishes shared facts; personas research only their angle.
2. Cap every sub-agent. Scout ≤ ~600 words; each persona ≤ ~700 words; Fact Checker ≤ ~700 words. Cite URLs. No preamble.
3. One round by default. Personas pre-empt the opposing side's strongest point. Rebuttal round ONLY for unresolved factual conflict. 1 to 2 agents max.
4. Orchestrator absorbs Neutral + Pragmatist + chairman in lite mode.
5. Parallelize. Multiple Task calls in one message per wave. Use agent summaries, not full transcripts.
6. Gather local context once. Read PRD/repo/files once; distil into Step-0 frame for every prompt.

Modes

• LITE (default): Scout → Advocate + Skeptic (parallel) → Fact Checker → orchestrator synthesis. ~4 sub-agents.
• FULL (/council full or user says "full council"): Scout → 5 personas + Fact Checker → rebuttal on contested points only → synthesis. ~2–3× lite tokens.

---

Procedure: LITE (default)

Step 0, Frame (orchestrator, inline). Read local context once. ≤6 bullets: decision, candidate options, hard constraints, single deciding metric. Paste into every Task prompt.

Step 1, Scout (1 Task). Invoke last30days when the decision depends on current social/web signal. Baseline: per-option price/spec/status, conflicts as ranges. ≤ ~600 words, URL per number, no opinion.

Step 2, Advocate + Skeptic (2 Tasks, parallel, one message). ≤5 lens-specific searches each:

• Advocate: strongest case for top 1–2 contenders; pre-empt Skeptic.
• Skeptic: adversarial downside; pre-empt Advocate.

Both: ≤ ~700 words, cite URLs, ranked top-N + factual claims for Fact Checker.

Step 3, Fact Checker (1 Task). deep-research Quick on claim list. ≤ ~700 words.

Step 4, Synthesize (orchestrator, inline). Fold in corrections. At most one targeted follow-up if a decisive number is still contested. Deliver:

• ranked list, #1 clearly marked, one deciding trade-off per option
• strongest preserved dissent
• "the condition that would change the answer"

Procedure: FULL (explicit request only)

Same Step 0 + Scout. Then Round 1 (5 parallel Tasks): Optimist, Pessimist, Neutral Analyst, Pragmatist + Fact Checker on strongest claims. Rebuttal on contested agents only. Synthesis with #1 marked, dissent preserved.

Output contract

Ranked list with #1 unambiguous; each entry one deciding line; main dissent visible; condition that would flip the verdict; honest "no option satisfies all constraints" when true. Chairman tone: decisive, evidence-bound.

Triggers (explicit invocation required)

Phrase	Mode
council this: …	LITE
/council …	LITE
/council full …	FULL
run the council, pressure-test this, stress-test this, war room this, debate this	LITE
"use council skill", @council	LITE unless user says "full"

Credits

• Methodology: Andrej Karpathy, LLM Council
• ECC council pattern: affaan-m/everything-claude-code
• Research companions: last30days, deep-research
• Token-efficient sub-agent pattern: Endokelp / CodingBrain (JARVIS brain-selection council, 2026-06)