devils-advocate
Systematically challenge current assumptions before major decisions. Counters confirmation bias, groupthink, and overconfidence.
haabe · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add haabe/mycelium --skill devils-advocateAbout this skill
name: devils-advocate description: "Systematically challenge current assumptions before major decisions. Counters confirmation bias, groupthink, and overconfidence." metadata: instruction_budget: "26" framework_dependency: "mycelium" framework_dependency_note: "This skill is designed to run within the Mycelium framework (https://github.com/haabe/mycelium). Standalone use will skip the canvas state, theory gates, and harness behavior the skill assumes. Install: /plugin install mycelium@haabe-mycelium."
Devil's Advocate
Run before every major diamond transition and architecture decision. Source: Kahneman, Shotton.
Technique 1: Pre-Mortem
Imagine it's 6 months from now and this decision FAILED spectacularly.
- What went wrong?
- What assumption was the weakest link?
- What signal did we ignore?
- Who was affected and how?
Technique 2: Assumption Reversal
For each key assumption:
- State the assumption explicitly
- Ask: "What if the OPPOSITE is true?"
- What evidence would support the opposite?
- Is there any evidence we've dismissed?
Technique 3: Red Team
Attack your own position:
- What would a competitor say about this approach?
- What would a skeptical user say?
- What would a security auditor find?
- What would an accessibility advocate flag?
10 Challenge Questions
- What are we most confident about? (That's where overconfidence hides)
- What evidence have we dismissed or downweighted?
- Are we anchored on our first idea? (Shotton - anchoring bias)
- Have we tested with users who DON'T match our ideal profile?
- What would make us abandon this direction entirely?
- Are we building for ourselves or for actual users?
- What's the simplest version that could validate/invalidate this?
- What have we NOT measured that we should?
- If we had to start over, would we make the same choice?
- Who disagrees with us and what's their strongest argument?
Technique 4: Attribution-vs-Consistency Check
Per anti-pattern Consistency-as-Evidence (#7) — graduated 2026-05-09 from a recurring failure where causal chains were built from observational consistency rather than verified attribution.
For each piece of evidence supporting the current claim, label it:
- Cleanly-attributed: the cause was demonstrably driving the effect (the action was Mycelium-specific, the variable was isolated, the alternative explanations were ruled out).
- Consistency-only: the data is compatible with the hypothesis but doesn't isolate the cause (the user reported X in a context where Y was also true; the trend matches the prediction but matches three other predictions equally well).
- Unrelated: the evidence is a different question entirely; don't include it in the chain.
If ≥1 link in a chain is consistency-only, mark the chain provisional and explicitly identify the missing attribution evidence. If N=1, do not publish a structural conclusion (e.g., "this generalizes to all users") until N≥2 with attribution. Apply this check to your own analysis pre-publish, not after the user catches it.
Technique 5: Ambient triggering on assertion-shaped patterns
Per the bias cluster (corrections.md TL;DR — L5 sycophancy, eval overfitting, sharper-framing anchoring; common root: "agent prefers what feels right over what evidence supports"):
Beyond formal diamond-transition use, run a fast self-check whenever you write text containing structural-claim shapes:
- "X causes Y"
- "this means Z"
- "the framework needs..." / "the user needs..."
- "the right answer is..."
- "this generalizes to..." / "this applies broadly..."
For each, ask: what specific evidence supports this claim, and does any of it merely support it by consistency rather than attribution? (Technique 4). If you can't name cleanly-attributed evidence for the claim, downgrade it: from assertion to hypothesis, from "X causes Y" to "X is consistent with Y; attribution evidence pending."
This converts the framework's own anti-bias discipline into a per-publish self-check, not just a per-decision ceremony. Graduated 2026-05-09 from corrections.md TL;DR open candidate.
MEASURED FIRE RATE, 2026-08-16: ZERO — and this note stays until that changes. In a dogfood session where twelve agent claims were withdrawn or corrected, this trigger fired on none of them. The operator caught six, explicit invocation of this skill caught two, late self-review caught three, one was caught proactively. A detector keyed on the five shapes above then scored 0 of 5 recall against the claims actually withdrawn, because those were universals and counts — "nobody left", "every instance", "eight statements" — not causal assertions. The five shapes named above are the wrong list for the failure population this technique exists to catch. Keep running it; do not treat it as coverage. Technique 7 is what measured well.
Technique 7: Blind adversarial pass — the one that measured well
USE THIS BEFORE A CLAIM LANDS ON A DURABLE SURFACE. Techniques 3 and 5 ask you to attack your own position. Measured against real withdrawn claims, self-attack found little and the ambient trigger found nothing. This found defects that had already survived self-review, four local gates and CI.
PREREQUISITE, AND CHECK IT BEFORE YOU START. This technique REQUIRES the ability to spawn a subagent. If subagent spawning is unavailable — the host forbids it, the user has asked that agents not be spawned unprompted, or the runtime has no such tool — then the technique with the best measured result is not available to you, and the fallback is self-attack, which this skill's own measurements describe as weaker. In that case you MUST (a) say so in the output, naming this technique as unavailable rather than silently substituting Technique 3 or 5, and (b) offer to run it if the user is willing to permit one subagent. Reported by a plugin consumer 2026-08-31, where a session-level no-subagent instruction silently downgraded the pass and not
Read the complete source ↗