Steelman

Anti-sycophancy. LLMs agree by default. This skill forces opposite case.

When invoked

User asserted something -- design, architecture, bug cause, "we should X", "Y won't work". Build strongest possible case against it before agreeing.

Procedure

Step 1: Identify claim

Restate user claim in one sentence. Flag type:

• Factual (verifiable: grep, docs, run) -> verify first
• Causal ("X breaks because Y") -> test the mechanism
• Architectural ("pattern Z won't scale") -> explore existing usage
• Preference/goal/scope -> decline steelman. Log noise. Return.

Preference/goal is user's call. No steelman.