devil-advocatelisted

# Devil's Advocate One fresh-context adversary whose only job is to prove the decision wrong. You judge. No consensus, no balance, no "on the other hand". A forced verdict at the end. **Core principle:** an argument you have been helping build cannot be attacked from inside the same context. The Advocate must start cold. ## When to use - Any decision that is expensive to reverse: rewrites, budget moves, hires, launches, migrations, "I'm presenting this tomorrow" - A conclusion drawn from data that fits a story everyone already believes - The user asks for pushback, holes, a stress test, or a red team Don't use for: factual questions, obvious execution tasks, code review (use a reviewer), or choosing between many open options (that is an LLM Council problem, see the last section). ## Workflow ### 1. Write the claim (you, in context) Reduce the decision to this block. If you cannot fill every line, the decision is not ready to be attacked. ``` CLAIM: <one sentence, falsifiable: "X will cause Y"> EVIDENCE FOR: <the 2–4 facts the claim actually rests on> COST IF WRONG: <what is lost — money, months, credibility> LOAD-BEARING ASSUMPTION: <the one thing that, if false, kills the claim> ``` Write the CLAIM exactly as big as the evidence supports. The Advocate attacks the sentence as written: "will remove most checkout errors" gets killed where "will remove the 2% of errors that are 503s" would not. Bundled actions ("rotate the key and scrub history") get attacked at their