Devils Advocate

Workflow

1. Understand the claim — Read the paper/argument being evaluated
2. Generate competing hypotheses — If evaluating a research question or design, load references/competing-hypotheses.md and generate 3-5 rival explanations before critiquing
3. Run the debate — Use the multi-turn debate protocol below (default) or single-shot mode for quick checks
4. Deliver the verdict — Synthesize surviving critiques with severity ratings

Multi-Turn Debate Protocol (Default)

Inspired by the simulated scientific debates in Google's AI Co-Scientist. A one-shot critique is easy for an LLM to produce but often superficial. Multi-turn debates force each critique to survive a defense, filtering out weak objections and sharpening the strong ones.

Round 1: Adversarial Critic

Adopt the persona of a hostile but competent reviewer. Challenge on:

1. Theoretical foundations — Are the assumptions justified?
2. Methodology — Limitations? Alternative approaches?
3. Data — Selection bias? Measurement issues? External validity?
4. Causal claims — Alternative explanations? Confounders?
5. Contribution — Novel enough? Does it matter?

Produce numbered critiques (aim for 5-8), each with a concrete statement of the problem.

Round 2: Defense

Switch persona to the paper's author. For each numbered critique, provide the strongest possible defense:

• Cite evidence from the paper that addresses the concern
• Explain design choices that mitigate the issue
• Acknowledge limitations honestly where the defense is weak
• Propose concrete fixes where the critique has merit

Round 3: Adjudication

Switch to an impartial senior reviewer. For each critique-defense pair, rule:

• Critique stands — the defense is insufficient; this is a real weakness
• Critique partially addressed — defense has merit but issue remains
• Critique resolved — the defense adequately addresses the concern

Final Synthesis

Produce a structured report with only the surviving critiques (stands + partially addressed), ranked by severity:

## Devil's Advocate Report

### Critical (must fix before submission)
1. [Critique] — [Why the defense failed] — [Suggested fix]

### Major (reviewers will likely raise)
2. [Critique] — [What remains after defense] — [Suggested fix]

### Minor (worth acknowledging)
3. [Critique] — [Residual concern] — [How to preempt]

### Dismissed
- [Critiques that were resolved in Round 2, listed briefly for transparency]

Single-Shot Mode

For quick checks (e.g., "just poke holes in this argument"), skip the multi-turn protocol and produce a direct critique. Use when the user says "quick", "just challenge this", or the input is a paragraph rather than a full paper.

Example Use

"Play devil's advocate on my research paper about preference drift - specifically challenge my identification strategy and the assumptions about utility functions."

Council Mode (Optional)

For the highest-stakes arguments, run the devil's advocate debate across multiple LLM providers. Different models have genuinely different reasoning patterns — a critique that Claude finds weak, GPT may find devastating, and vice versa. This produces adversarial tension that a single model cannot replicate internally.

Trigger: "Council devil's advocate" or "thorough challenge"

How it works:

1. Each model independently plays Adversarial Critic (Round 1) using the same paper/argument
2. Cross-review: each model evaluates the others' critiques — identifying which challenges are strongest
3. Chairman synthesis: produces a single report with surviving critiques ranked by cross-model agreement

Invocation (CLI backend):

cd packages/cli-council
uv run python -m cli_council \
    --prompt-file /tmp/devils-advocate-prompt.txt \
    --context-file /tmp/paper-content.txt \
    --output-md /tmp/devils-advocate-council.md \
    --chairman claude \
    --timeout 180

See skills/shared/council-protocol.md for the full orchestration protocol.

Value: High — the multi-turn debate protocol (Round 1→2→3) becomes genuinely adversarial when different models play different roles. A critique that survives cross-model scrutiny is almost certainly a real weakness.

Cross-References