agentleFS
Sign inStart free
FuDesign2008FuDesign2008/open-skills/skills/multi-agent-debate/SKILL.md

multi-agent-debate

Skill6 starsChanged 6 days agoMITView on GitHub

Stress-test a crash analysis, hypothesis, or technical conclusion by launching three adversarial agents in parallel — one defends, one attacks, one hunts new evidence — and resolving disputes with physical proof instead of argument. Always use this skill when the analysis has 2+ competing explanations that both seem plausible, when a conclusion relies on a single log sample or inferred chain longer than 3 steps, or when someone says \"are you sure?\", \"that doesn't sound right\", or challenges an assumption. Triggers: 辩论、求真、挑战假设、质疑分析、multi-agent-debate, \"debate this\", \"challenge my analysis\", \"find holes in\", \"I'm not sure this is right\", \"verify this\".

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Write your hypothesis in one paragraph

    line 14
  2. Distinctive:

    No bare claims in conclusions.

    line 72

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Multi-Agent Debate

Core rule: Reasoning disputes cannot be resolved by more reasoning. Only evidence arbitrates.

Quick Start

  1. Write your hypothesis in one paragraph
  2. List existing evidence (each with source: file, line, or command)
  3. Identify 2–3 weakest links in your analysis
  4. Fill in the three prompts below and launch in parallel
task(subagent_type="explore",  run_in_background=true, prompt="[DEFENDER]...")
task(subagent_type="oracle",   run_in_background=true, prompt="[CHALLENGER]...")
task(subagent_type="explore",  run_in_background=true, prompt="[EVIDENCE HUNTER]...")

Why separate prompts? Each agent must be blind to the others' goals — if the Challenger sees the Defender's framing, it anchors on the same assumptions it's supposed to attack. Fill in each prompt from REFERENCE.md independently.

Minimal prompt structure (fill in <...>):

DEFENDER:   "You are the DEFENDER for: <hypothesis>. Find direct physical evidence
             (file:line, command output, code) for each claim. List the 3 weakest
             points and collect pre-emptive counter-evidence."

CHALLENGER: "You are the CHALLENGER for: <hypothesis>. Find holes: alternative
             explanations, single-sample claims, inferences used as facts. For
             every alternative you propose, provide how to falsify it."

HUNTER:     "Find facts only — no inference. Answer: <2-3 specific questions from
             disputed claims>. Every answer must be a command output or code verbatim.
             Tag each: [SUPPORTS] / [CHALLENGES] / [NEUTRAL]."

Three Roles

RoleGoalAgent Type
DefenderFind strongest evidence for the hypothesisexplore
ChallengerFind holes, alternative explanations, single-sample claimsoracle
Evidence HunterFind new facts only — no inference allowedexplore

5-Step Flow

  1. Prepare — hypothesis + evidence list + 2-3 weakest links
  2. Launch — all three agents in parallel (run_in_background=true)
  3. Triage — split results into: agreed facts / disputed claims / new evidence
  4. Arbitrate — for each dispute, find a command/file/code that directly resolves it; if only reasoning exists, mark [UNRESOLVED]
  5. Conclude — write only what has physical evidence; disputed items stay pending

Conclusion Format

### Confirmed Facts [FACT:source]
### Reasonable Inferences [INFERENCE]
### Unresolved Disputes — verification method needed
### Corrected Mistakes — original → corrected | evidence

Evidence Tags

[FACT:file:line] · [FACT:cmd] · [FACT:N samples] · [INFERENCE] · [SPECULATION] · [UNRESOLVED]

No bare claims in conclusions. Every statement needs a tag.

See REFERENCE.md for full prompt templates with examples, common traps, arbitration decision tree, and evidence priority rules.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free