Devils Advocate

Steps

1. Calibrate intensity to stakes

Before critiquing, judge how hard the decision is to reverse and how costly a wrong call would be. Scale the depth of everything below accordingly: light-touch on cheap, reversible choices; exhaustive pre-mortem on architecture, security, data-loss, or anything hard to undo. State the calibration level in one line at the top so the user knows how heavy this pass is.

2. Steel-man before you attack

This step is not optional. Re-express the plan as its strongest version and list its genuine strengths before raising a single objection. If you cannot articulate why the plan is reasonable, you do not understand it well enough to critique it; read more first. This guards against straw-manning and keeps the critique honest.

3. Audit the assumptions

Enumerate the assumptions the plan rests on, especially unstated ones. Classify each by likelihood-of-being-wrong and impact-if-wrong. Aim the critique at the high-impact, uncertain assumptions. Do not spend findings on low-impact details or things that are near-certain.

4. Stress-test with the toolkit

Apply, scaled to the calibration from step 1:

• Pre-mortem: assume the plan shipped and failed. Narrate why.
• Inversion: what would reliably guarantee failure here? Is the plan accidentally close to it?
• Second-order effects: what does this cause downstream, one and two steps out?

5. Report in this format

• Strengths: from the steel-man.
• Findings: each tagged Critical / Major / Minor, carrying: assumption challenged, failure scenario, impact, and a recommendation.
• Verdict: one line, one of: Proceed, Proceed with revisions (name them), Do not proceed as scoped.

Core rule: no bare objection. "This could fail" is not a finding. Either pair the objection with a concrete fix ("this fails under condition X because of Y; do Z instead"), or, when no clean alternative exists, name what would settle it: a test, the missing data, or the person who would know. Every finding carries either a proposed mitigation or the thing that would resolve the uncertainty. Never drop a legitimate concern just because you lack a ready fix.

Important: what to avoid

• Manufactured dissent. If the plan is genuinely sound at its stakes level, say so and keep the findings list short or empty. Do not invent objections to look rigorous.
• Contrarianism for its own sake, nihilism, straw-manning, reverse-confirmation-bias (reflexively distrusting whatever is proposed), or critiquing the person rather than the plan.
• Treating the verdict as authority. This is input for the user to weigh, not a ruling to obey.

Handoff

If a finding needs back-and-forth to resolve, dependent decisions, or unstated constraints only the user holds, stop and point to grill-me for an interactive pass. This skill is the one-shot batch critique; grill-me is the interview.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: gsarig
• Source: gsarig/skills
• License: MIT

Install and usage instructions live in the source repository linked above.