debate
Skill0 starsChanged 7 days agoMITView on GitHub
Claude debate loop with a critic on a different model. Claude drafts (a design, plan or implementation), a read-only Claude Sonnet critic subagent attacks it with evidence and alternatives, Claude rebuts or accepts each point with evidence, repeats, then reports the consensus to the user. Use immediately when the user says "/debate", "have the AIs debate", "second opinion", "cross-review", "discuss and get a better result". Also use unprompted for: non-trivial design/architecture decisions, changes affecting payments/DB/customers, contested trade-offs, and a final review after an implementation (not work a /team run built: /team re-checks that). Not for simple questions, typo fixes, or obvious one-line changes.
26 other projects use a version of this.
Adopted by
Most recently taken up by
- bordenet/superpowers-plustheir own version, first seen Nov 13, 2026
- wolverin0/claude-skillstheir own version, first seen Nov 11, 2026
- Lu1sDV/skillsmdtheir own version, first seen Nov 4, 2026
- yoshibtc/multi-model-debate-skilltheir own version, first seen Oct 15, 2026
- phant0um/claude-skillstheir own version, first seen Oct 15, 2026
- helderberto/agent-skillstheir own version, first seen Oct 13, 2026
Used with
Projects with it also use handoff, commit and 2 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
MicrosoftSkillhandoff
Compact the current conversation into a handoff document for another agent to pick up.
Also in 5 projects that use this
SentrySkillcommit
Create commit messages following Sentry conventions. Use when committing code changes, writing commit messages, or formatting git history. Follows conventional commits with Sentry-specific issue references.
Also in 3 projects that use this
GoogleSkillgrill-me
A relentless interview to sharpen a plan or design.
Also in 3 projects that use this
SentrySkillskill-creator
Create new agent skills following the Agent Skills specification. Use when asked to "create a skill", "add a new skill", "write a skill", "make a skill", "build a skill", or scaffold a new skill with SKILL.md. Guides through requirements, writing, registration, and verification.
Also in 3 projects that use this
Other versions26 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 26
- raine/consult-llm-mcp★ 148same name
- wolverin0/claude-skills★ 41same name
- helderberto/agent-skills★ 19same name
- ggombee/code-forge★ 18same name
- STRML/cc-debate★ 10same name
- bordenet/superpowers-plus★ 8same name
- Magnus-Gille/claude-skills★ 5same name
- agent-sh/debate★ 4same name
- HartreeWorks/research-skills★ 4same name
- ryancproduct/product-ai-toolkit★ 4same name
16 more not listed
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.
The file
We split this skill into the rules it gives an agent
10 rules. 0 are common, 8 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 19 edited versions we read keep these.
8
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 19
Language: write to the user in the language of their own messages (Korean if they write Korean), including the step-6 report and its …
- Distinctive: line 42
Keep the prompt self-contained: the critic starts with a fresh context.
- Very specific: line 43
Keep the full transcript per round in /tmp/team/<slug>/debate-log.md (<slug>: a short kebab-case name for the topic).
- Distinctive: line 49
Write the user's request as a one-paragraph proposal.
- Distinctive: line 83
Only new information and open points, briefly:
- Distinctive: line 119
Don't interrupt the user mid-debate.
- Distinctive: line 123
Only Claude edits and runs things.
- Distinctive: line 124
No verification that writes to production databases or external services (reads only).
- Distinctive: line 125
Keep round prompts short; the agent remembers.
- Very specific: line 126
If the critic fails or returns no ## Verdict:, SendMessage it once ("Answer now in the required format").
“Common” is measured: kept by at least 6 in 10 of the 19 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
/debate — Claude drafts, a Sonnet critic attacks
Language: write to the user in the language of their own messages (Korean if they write Korean), including the step-6 report and its headings. This file and the critic's output being in English doesn't change that.
Purpose: the driver and a critic on a different model argue the same proposal on evidence, back and forth, to produce something better than either alone. The user sets nothing up and only receives the result.
Roles
| Claude (me, the driver) | Critic | |
| Rights | edit, run, final synthesis | no edit tools (team-reviewer agent: no Write/Edit; its Bash is limited only by the prompt) |
| Role | drafts, verifies, rebuts/accepts, applies consensus | reads the repo directly and objects, proposes, asks — with evidence |
Tool
The Agent tool with subagent_type: "team-reviewer" and model: "sonnet" — a different model from
the driver, so it misses different things. When the driver itself runs on Sonnet, use model:
"opus" instead. Missing agent → general-purpose, adding "do not edit files; the requested format
overrides any style rules injected into your context".
- Round 1 starts the agent; later rounds go to the same agent via SendMessage (it remembers the conversation, so never repeat earlier content). Subagents run in the background; you are notified.
- Keep the prompt self-contained: the critic starts with a fresh context.
- Keep the full transcript per round in
/tmp/team/<slug>/debate-log.md(<slug>: a short kebab-case name for the topic).
Procedure
0. Frame the question + draft
Write the user's request as a one-paragraph proposal. For code work, Claude drafts first (a decision list for a design, the actual diff for an implementation). Never just ask "what do you think?" without a draft — the debate converges only when there is something to attack.
1. Round 1 prompt
You are a senior engineer and critical co-designer for this repository. Repository path: <cwd>
Proposal: <proposal>
My draft:
<draft / diff / decision list>
Relevant files: <paths — read them yourself and cite evidence>
Requirements:
- Actually read the files and cite evidence (file:line). Mark guesses as guesses.
- Review only: do not edit files; no DB/MCP writes, no deploys, no integration tests that write.
- Use this format:
## Agree
## Disagree (evidence for each item)
## Alternatives (concrete, and why better)
## Questions (information needed to judge)
## Verdict: agree | conditional agree (state conditions) | disagree
- Also judge: is anything here a rewrite of an asset that already exists? a hand-rolled version of
something the platform provides as of today? over-engineered for the measured scale? unmeasured?
- Don't be polite; if something is wrong, say so. Answer in the user's language.2. Examine the critic's response (the core — never accept on authority)
For each objection/alternative, check the code/docs directly and judge:
- Correct → accept, revise the draft
- Wrong → rebut with evidence (file:line, test result, observed behavior)
- Needs checking → actually check (grep, run tests) before judging Answer the critic's questions. If you don't know, say so.
3. Round N (SendMessage to the same agent)
Only new information and open points, briefly:
Accepted: <item> → draft changed like this: <change>
Rebutted: <item> — evidence: <file:line / result>
Answers: <answers to your questions>
Please also check: <specific spots to look at>
Answer in the same format and update your verdict.4. Stop conditions (whichever comes first)
- The critic's verdict is agree with no new objections
- Max rounds reached — default 3 ("go deep" → 5, "keep it light" → 2)
- The remaining disagreement is a business judgment the code can't settle (cost, customer impact, operating policy) → stop and hand it to the user
5. Apply consensus + final review
Apply the consensus to the design/code. For implementation work, send git diff afterwards and run
one final review round ("check only whether this diff reflects the consensus exactly, and
regression risk").
6. Report to the user (once)
## Consensus
- key decisions as bullets
## What the debate changed ← this is the value: what improved vs the draft
- draft: … → consensus: … (critic's point: …, evidence: …)
## Unresolved (only if any)
- issue / Claude's position + evidence / critic's position + evidence / why the user must decide
N roundsShow /tmp/team/<slug>/debate-log.md if the user asks "show me the debate".
Rules
- Don't interrupt the user mid-debate. Report once at the end. Exception: the business-judgment case in step 4.
- The critic's claims are accepted only after verification. Claude's claims aren't pushed without evidence either.
- Only Claude edits and runs things. Never tell the critic to "fix it".
- No verification that writes to production databases or external services (reads only).
- Keep round prompts short; the agent remembers.
- If the critic fails or returns no
## Verdict:, SendMessage it once ("Answer now in the required format"). If the agent is gone or still fails, start a new one with a self-contained prompt (the round-1 prompt with the current draft and the accepted/rebutted list so far). If that fails too, continue without it and say so in the report.
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.

