Most council skills give you opinions.
This one gives you a ranked verdict, checked claims, and the dissent you would have missed.
Install · Demo · Usage · Should I Build
CouncilDemo.mp4
90-second demo (real lite run, Mac Studio vs used 4090 PC)
Lite mode by default (~4 sub-agents). Scout pulls recent signal via last30days. Fact Checker verifies numbers via deep-research. You leave with #1 marked, the strongest counterargument still in the room, and the one condition that would flip the call.
Works in Cursor and Claude Code. No server. Inference is the assistant you already pay for.
npx skills add Endokelp/DecisionCouncil --skill council -g -yCompanion research skills in one shot:
Windows (PowerShell):
git clone https://github.com/Endokelp/DecisionCouncil.git $env:TEMP\DecisionCouncil
& "$env:TEMP\DecisionCouncil\install.ps1"macOS / Linux:
git clone https://github.com/Endokelp/DecisionCouncil.git /tmp/DecisionCouncil
/tmp/DecisionCouncil/install.shRestart your editor after install.
The skill only runs when you name it (disable-model-invocation: true). Say one of:
| You type | Mode |
|---|---|
council this: Mac Studio vs DGX Spark under $5k |
Lite (default) |
/council Which vendor for a 3-year lock-in? |
Lite |
/council full Should we pivot pricing? |
Full (~2 to 3x tokens) |
pressure-test this, war room this, debate this |
Lite |
Good questions: vendor lock-in, hardware purchases, architecture forks, strategy pivots, anything expensive to get wrong.
Skip it: trivia, cheap reversible choices, pure writing, code review (use a reviewer skill).
## Council: Local AI brain under $5k
1. **Mac Studio M4 Max 128GB** (WINNER)
Trade-off: swap-on-demand for 70B, not resident dual-30B parallel.
2. DGX Spark 128GB
Trade-off: dense 70B decode ~5 tok/s; thermal reports still messy.
3. Dual used RTX 3090 build
Trade-off: best tok/s per dollar, wrong for a quiet 24/7 desk box.
**Strongest dissent:** Skeptic argues Spark CUDA ecosystem wins if fine-tuning becomes primary, not 20% side work.
**Condition that would change the answer:** verified firmware lifts Spark dense-70B above ~15 tok/s AND fixes sustained thermal shutdowns.| affaan-m/council | karpathy/llm-council | aiwithremy/llm-council | DecisionCouncil | |
|---|---|---|---|---|
| What it is | Opinion council in a skill pack | Multi-model web app | 5-advisor Claude skill | Standalone evidence council |
| Research | You bring context | None built-in | None built-in | last30days Scout + deep-research Fact Checker |
| Output | Compact verdict + dissent | Chairman synthesis | Chairman + peer review | Ranked list, #1 marked + falsifier |
| Token model | 3 sub-agents, one round | Full pipeline always | 5 advisors + review | Lite default, Full optional |
| Best for | Fast ambiguous tradeoffs | Compare models | Business/creator calls | High-stakes, evidence-bound forks |
Karpathy compares models. The others convene opinions. This one is for when the decision has a price tag.
Install lands at ~/.cursor/skills/council/. Invoke with council this: or /council.
Sub-agents use the Task tool (generalPurpose, run_in_background: false). Spawn Scout, then Advocate + Skeptic in parallel, then Fact Checker. Chairman synthesis stays in the main session.
Details: skills/council/SKILL.md
Also installs to ~/.claude/skills/council/. Use /council or /council full.
DecisionCouncil/
├── CouncilDemo.mp4
├── council-hero.png
├── logo.svg
├── AGENTS.md
├── skills/council/SKILL.md
├── install.ps1
├── install.sh
└── README.md
Demand first, then the fork: Should I Build.
New to agent skills? Only Skill You Need.
- Methodology: Andrej Karpathy, LLM Council
- ECC council pattern: affaan-m/everything-claude-code
- Community skill: aiwithremy/claude-skills-llm-council
- Research companions: last30days, deep-research
- Built by Endokelp. Dogfooded on a real hardware council (JARVIS brain selection, 2026-06).
MIT. See LICENSE.