| name | multi-ai-council |
| description | Use when the user wants a hard question, proposal, design, review, tradeoff analysis, or difficult debugging problem answered by a council of Claude, Codex, and Gemini through this repo's local multi-ai CLI. Trigger for requests like "ask multiple AIs", "use the council", "get another model's take", "high-confidence review", "difficult question", or "compare top models on this". |
Multi-AI Council
Use this skill when one model's answer is not enough and the user would benefit from a multi-model pass.
This repo already provides the real execution entrypoint:
The default base path is the current working directory. That means the council naturally runs against the repo or folder you invoke it from.
When to use it
Prefer this skill for:
- difficult architecture or design tradeoffs
- proposal/spec review
- high-stakes correctness or soundness questions
- bug analysis where multiple perspectives are useful
- "ask Claude, Codex, and Gemini" style requests
Do not use it for:
- trivial factual questions
- simple rewrites
- tasks where the overhead of a council would dominate the value
First decision: run mode vs review mode
The most important choice is the worker mode, not the preset.
Use --mode run when you want open-ended answers:
- architecture questions
- design tradeoffs
- brainstorming
- explanation or synthesis
- "what is the best approach?" style prompts
In run mode, providers answer freely in text and the leader synthesizes.
Use --mode review when you want concrete issues:
- code review
- spec review
- proposal review
- soundness/risk review
- "find problems, fake progress, missing invariants, or correctness holes"
In review mode, providers return structured findings, the council merges and deduplicates them, and the leader synthesizes from that findings set.
Short mental model:
--mode run = ask several models, then synthesize
--mode review = ask several models for findings, merge them, optionally debate them, then synthesize
Debate
If you specifically want debate, use --mode review --debate.
Important current behavior:
- real debate runs in
review mode
- do not rely on
--debate in run mode for the same behavior
So:
- open-ended hard question:
--mode run
- review/findings/risk pass:
--mode review
- review with actual debate:
--mode review --debate
Presets
fast: quick pass; best for exploratory --mode run
default: normal general-purpose pass
review: good default preset for --mode review
rigor: use for the hardest review questions when extra rigor is worth the time
Presets tune the pass. Modes change the worker contract. Do not confuse them.
Workflow
- Choose the mode first:
--mode run for open-ended answers
--mode review for findings
- Choose the preset that matches the depth you want.
- Run
multi-ai from the current workspace:
multi-ai run --mode review --preset review "Review this proposal for correctness and soundness."
- Read the JSON or final text output.
- Summarize the leader answer for the user.
- If useful, call out disagreements, merged findings, or notable worker failures.
Notes
- The council uses the local
multi-ai CLI, not direct API calls.
- The final answer comes from the configured leader after the worker phase.
- The CLI entrypoint is still
multi-ai run; the choice is --mode run vs --mode review.
- If the user wants raw worker outputs too, prefer
--json and inspect the providers object.
Examples
General hard question:
multi-ai run --mode run --preset default "What is the best architecture for this feature?"
Spec review:
multi-ai run --mode review --preset review "Review tmp/spec.md for correctness, soundness, and missing edge cases."
Review with actual debate:
multi-ai run --mode review --preset rigor --debate --json "Review this design in detail."