Devils Advocate

Devil's Advocate (Full Analysis)

Context: $ARGUMENTS

Run a complete adversarial analysis by orchestrating all three sub-skills in sequence. Each builds on the shared argument graph.

Workflow

Phase 1 — Steelman + Contrary Case

Read ${CLAUDE_PLUGIN_ROOT}/references/steelmanning.md and ${CLAUDE_PLUGIN_ROOT}/references/dissoi-logoi.md.

Reset and initialize:

bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" reset
bun "${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts" reset
bun "${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts" init "<subject>" --total 7

If the user asked to watch the analysis, start the live transcript in the background first — never otherwise, since it opens a browser:

bun "${CLAUDE_PLUGIN_ROOT}/scripts/transcript.ts" serve &

Build the strongest version of the argument, then argue the complete opposite with equal conviction. Register both in the graph as you go — a claim that never reaches the graph cannot be defended in Phase 3.

bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" add "<claim-id>" claim --obs "<fact>" --agent advocate
bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" add "<assumption-id>" assumption --obs "<fact>" --agent advocate
bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" relate "<claim-id>" assumes "<assumption-id>" --agent advocate
bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" add "<evidence-id>" evidence --obs "<source>" --agent advocate
bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" relate "<evidence-id>" supports "<claim-id>" --agent advocate

Ground every assumption you can. An assumption with no supports edge demotes the claim resting on it.

bun "${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts" think 1 "<steelman>"
bun "${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts" think 2 "<contrary case>"

Present both to the user before continuing.

Phase 2 — Crucible

Read ${CLAUDE_PLUGIN_ROOT}/references/socratic-method.md, ${CLAUDE_PLUGIN_ROOT}/references/crucible.md, and ${CLAUDE_PLUGIN_ROOT}/references/red-team.md.

Ask 3-5 Socratic questions targeting specific assumptions. Run structural analysis:

bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" analyze

Use the gaps to drive the hostile review. Register every weakness as a node that attacks what it damages — an unattached weakness changes no verdict.

bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" add "<weakness-id>" weakness --obs "<what breaks>" --agent adversary
bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" relate "<weakness-id>" undermines "<claim-id>" --agent adversary
bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" relate "<claim-a>" contradicts "<claim-b>" --agent adversary

A failure mode attacks — use undermines. if_fails is different: it means "if the source fails, the target fails", so it links two claims where one carries the other. Do not point if_fails from a weakness at a claim; that makes the claim depend on the weakness holding, which inverts the result.

bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" relate "<load-bearing-claim>" if_fails "<dependent-claim>" --agent adversary

Then answer your own attacks honestly. A weakness with a surviving counter or mitigation stops counting against the claim — that is the mechanism, not a courtesy:

bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" add "<counter-id>" counter --obs "<why the attack fails>" --agent advocate
bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" relate "<counter-id>" undermines "<weakness-id>" --agent advocate
bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" relate "<mitigation-id>" mitigates "<weakness-id>" --agent advocate

Identify 3-5 failure modes. If multiple independent failure modes exist, branch:

bun "${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts" think 3 "<socratic questions>"
bun "${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts" think 4 "<crucible findings>"
bun "${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts" think 5 "<red team findings>"

REVISION CHECK: If the crucible reveals the steelman was too generous, revise:

bun "${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts" revise 1 "<revised steelman>" --reason "<what changed>"

Phase 3 — Verdict

Read ${CLAUDE_PLUGIN_ROOT}/references/social-gadfly.md.

Name the one uncomfortable truth nobody wants to say. Then run the structural verdict:

bun "${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts" think 6 "<gadfly sting>"
bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" verdict
bun "${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts" think 7 "<verdict>"

Output format

Present the full analysis with clear sections:

1. The Strongest Case: (steelman)
2. The Strongest Case Against: (dissoi logoi)
3. Socratic Questions: (3-5 questions targeting assumptions)
4. Weakest Points: (crucible findings from structural analysis)
5. Failure Modes: (red team, with branches if multiple)
6. The Uncomfortable Truth: (gadfly sting)
7. Verdict: Survives / Does not survive / Needs more work

Optionally render the full chain and graph:

bun "${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts" render
bun "${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts" export

Resources

References (load the file matching the current phase):

• ${CLAUDE_PLUGIN_ROOT}/references/steelmanning.md — load in Phase 1.
• ${CLAUDE_PLUGIN_ROOT}/references/dissoi-logoi.md — load in Phase 1.
• ${CLAUDE_PLUGIN_ROOT}/references/socratic-method.md — load in Phase 2.
• ${CLAUDE_PLUGIN_ROOT}/references/crucible.md — load in Phase 2.
• ${CLAUDE_PLUGIN_ROOT}/references/red-team.md — load in Phase 2.
• ${CLAUDE_PLUGIN_ROOT}/references/social-gadfly.md — load in Phase 3.
• ${CLAUDE_PLUGIN_ROOT}/references/polemic.md — load when the gadfly sting needs rhetorical force.
• ${CLAUDE_PLUGIN_ROOT}/references/devils-advocate.md — load for Nemeth's warning about inauthentic dissent.
• ${CLAUDE_PLUGIN_ROOT}/references/roman-rota.md — load when multi-level appellate review structure is relevant.
• ${CLAUDE_PLUGIN_ROOT}/references/lawsuits-against-the-devil.md — load when the idea seems too absurd or sacred to challenge.
• ${CLAUDE_PLUGIN_ROOT}/references/argument-semantics.md — load always. How the verdict is computed: attacks from defeated nodes stop counting, so counters decide outcomes.

Scripts:

• ${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.ts — run throughout. analyze before Phase 2, verdict in Phase 3.
• ${CLAUDE_PLUGIN_ROOT}/scripts/thought-chain.ts — run at every phase. revise when later phases invalidate earlier ones. branch for multiple failure modes.

Gotchas

• Nemeth's warning is real. Inauthentic, going-through-the-motions dissent makes people MORE entrenched. Every objection must cite a specific mechanism of failure.
• Steelman FIRST, always. If you attack a weak version, the entire analysis is worthless.
• Run analyze before the crucible. The graph catches structural gaps prose misses.
• Revise when warranted. A linear pass that never revisits is weaker than one that evolves.
• The gadfly sting is not another weakness. It's the foundational uncomfortable truth.
• Graph entity names: short kebab-case. Observations carry the detail.