agentleFS
sungminpark-bizsungminpark-biz/tandem-skills/skills/debate/SKILL.md

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

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.
  • MicrosoftSkill

    handoff

    Compact the current conversation into a handoff document for another agent to pick up.

    Also in 5 projects that use this

  • SentrySkill

    commit

    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

  • GoogleSkill

    grill-me

    A relentless interview to sharpen a plan or design.

    Also in 3 projects that use this

  • SentrySkill

    skill-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

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
  1. Distinctive:

    Language: write to the user in the language of their own messages (Korean if they write Korean), including the step-6 report and its …

    line 19
  2. Distinctive:

    Keep the prompt self-contained: the critic starts with a fresh context.

    line 42
  3. Very specific:

    Keep the full transcript per round in /tmp/team/<slug>/debate-log.md (<slug>: a short kebab-case name for the topic).

    line 43
  4. Distinctive:

    Write the user's request as a one-paragraph proposal.

    line 49
  5. Distinctive:

    Only new information and open points, briefly:

    line 83
  6. Distinctive:

    Don't interrupt the user mid-debate.

    line 119
  7. Distinctive:

    Only Claude edits and runs things.

    line 123
  8. Distinctive:

    No verification that writes to production databases or external services (reads only).

    line 124
  9. Distinctive:

    Keep round prompts short; the agent remembers.

    line 125
  10. Very specific:

    If the critic fails or returns no ## Verdict:, SendMessage it once ("Answer now in the required format").

    line 126

“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.

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

/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
Rightsedit, run, final synthesisno edit tools (team-reviewer agent: no Write/Edit; its Bash is limited only by the prompt)
Roledrafts, verifies, rebuts/accepts, applies consensusreads 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 rounds

Show /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.

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.