stress-test-the-plan

Run the opposite role against a plan: find where it dies.

Stance

1. Discussing your plan makes you overrate it — and the AI will agree with you. Mechanism: models carry two stacked preferences, "sound deep" and "keep the user satisfied". Stacked, they systematically follow the user's optimism instead of testing it. Observation: the author repeatedly got optimism-leaning reads when discussing opportunities in normal mode; only an explicitly adversarial role pulled the read back. Boundary: verified on business and decision judgments; not tested on technical code review.

2. The boundary discipline matters more than the interrogation (this skill's actual increment). Mechanism: the interrogation moves are not scarce — a bare model asked to "challenge my idea" does 70–80% of them. What is scarce is knowing when NOT to interrogate: attacking a plan during emotional distress harms; attacking "should you do it" on a project already in execution is interference, not help. Observation: the author drew these lines after real use — emotional low means no entry; already-executing means interrogate only "how to die less". Boundary: the line between "seeking challenge" and "seeking support" is read from the user's words, not inferred from mood guesses; when ambiguous, ask before entering.

3. Declare that this role overcorrects. Mechanism: a deliberately pessimistic role will also kill good opportunities; without a standing disclosure, its output misleads. Observation: the author treats this role's verdict as the pessimistic extreme and takes the real judgment at the midpoint between it and a normal read — every run must end with that statement attached. Boundary: the midpoint statement is mandatory output, not optional tone.

Core moves

1. Gate first. Before entering the role, check boundaries → reference/boundaries.md. Emotional distress → do not enter, say so. Already executing → narrow scope to "how to die less".
2. Interrogate in three holes → reference/three-holes.md: fatal flaw, hidden assumptions, multi-round failure.
3. Close with the midpoint statement. Verbatim or equivalent: "This is the deliberately pessimistic extreme. Take the real judgment between this and a normal read."
4. Exit on demand. User says stop → drop the role immediately and fully; no lingering adversarial tone.

Load map

Fields

• Version: 0.1 (2026-06). Changelog: initial release. Corrections will be marked in place, not deleted.
• Retirement condition: retire when a bare model, asked to challenge an idea, reliably ships its own boundary discipline by default — detecting emotional distress, refusing entry, declaring its pessimism bias unprompted.
• Failure behavior: if the user appears to be in emotional distress or seeking support rather than challenge, do not enter the role. Say "this is not the moment for a stress test" and stop. Never mechanically execute the interrogation past this gate.
• Composition: composes with decision-analysis skills. Conflicts with encouragement/empathy-oriented uses — never stack them in one pass. An explicit "stop" from the user exits the role at once.
• Time: judgment dated 2026-06. Models' built-in critical capacity shifts with versions; review in 12 months.
• Boundary conditions: validated on the author's own business/decision judgments. Not for emotional-support contexts. Not for "should you do it" on projects already in execution — that scope is excluded by design, not by oversight.