adversarial-debate
Skill65 starsChanged 7 months agoMITView on GitHub
Run an Advocate/Critic/Judge debate over any decision with real tradeoffs. Use when the user asks whether to merge a PR, pick an architecture, hire a candidate, ship a feature, or make any judgment where both sides have valid points and a single-model answer would be too confident.
2 other projects use a version of this.
Adopted by
Most recently taken up by
- kuangketongxue/kuang-skillstheir own version, first seen Sep 9, 2026
- alex-jb/claude-debateword for word, first seen Apr 17, 2026
Other versions2 projects use a version of this.
Word for word · 1
- alex-jb/claude-debate★ 2100% same
Edited · 0
None found.
Wrote their own version · 1
- kuangketongxue/kuang-skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository4 other files it gives its agents.
- orallexa-ui/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- orallexa-ui/CLAUDE.mdCLAUDE.md
- tier-routerSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
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
- Distinctive: line 19
Should we merge this PR given tests, blast radius, author tenure?
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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.
Files
Adversarial Debate
Three-stage debate pattern for high-stakes decisions:
- Advocate (Haiku) — strongest case FOR the proposal
- Critic (Haiku) — strongest case AGAINST, rebutting the advocate
- Judge (Sonnet) — synthesize and return a structured verdict
Cost ~$0.003 per debate. Takes ~10-15 seconds.
When to use
Good fits:
- Should we merge this PR given tests, blast radius, author tenure?
- Postgres vs DynamoDB for this workload?
- Hire or pass after the last round?
- Ship this feature now or wait a week?
- Trading / investment: enter or stay out?
- Any decision with a defensible case on both sides
Bad fits:
- Questions with one right answer (use a regular call)
- Pure factual extraction (use a cheap single call)
- Open-ended brainstorming (debate narrows, not expands)
- Low-stakes decisions where $0.003 + 15s isn't worth it
Usage
from anthropic import Anthropic
from claude_debate import run_debate, DebateConfig
verdict = run_debate(
proposal="Merge PR #42: adds Redis cache to auth middleware",
client=Anthropic(),
context={
"diff_size": 340,
"coverage": "85%",
"author_tenure": "3 weeks",
"prod_qps": 2400,
},
config=DebateConfig(
decision_options=("APPROVE", "REQUEST_CHANGES", "DEFER"),
criteria=["correctness", "rollback cost", "perf under load"],
),
)
print(verdict.decision) # "APPROVE"
print(verdict.confidence) # 72.0
print(verdict.reasoning_summary) # one sentence
print(verdict.strongest_advocate_point) # what was most convincing for
print(verdict.strongest_critic_point) # what was most convincing againstDesign choices worth knowing
- Critic sees the advocate's case — so the critic's points are specific rebuttals, not independent arguments. This is what makes the debate adversarial instead of two parallel monologues.
- Judge is explicitly asked for its tipping condition — "what evidence would flip you?" — so the verdict is falsifiable, not just assertive.
decision_optionsare bounded — the judge can't invent a new option. Forces a commit.- Confidence is calibrated in the prompt — 80+ = would bet on it, 50 = genuinely uncertain. Not a vibe.
- Cost asymmetry matches reasoning asymmetry — advocate and critic are structured argument generation (Haiku is fine); judge is reasoning over conflicting evidence (Sonnet).
Anti-patterns
- Don't run debates on questions where one side is obviously right. The advocate/critic will manufacture weak arguments and the verdict is noise.
- Don't skip the
contextdict on non-trivial decisions. Without facts, both sides speculate and the judge has nothing to weigh. - Don't treat the verdict as binding. It's a structured second opinion — use it to pressure-test your own instinct, not replace it.
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
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.

