Rubber Duck
Adversarial "rubber duck" review that turns explaining-out-loud into a hallucination...
Install
npx skills add kirodotdev/KiroCrew --skill rubber-duck -g -a claude-code -y- Repository
- kirodotdev/KiroCrew4.7k
- Category
- Design
- Updated
- 3d ago
- License
- Apache-2.0
- Context cost
- Mediumabout 2,405 tokens
- Runs scripts
- No, instructions only
What it does
Adversarial "rubber duck" review that turns explaining-out-loud into a hallucination check. The main session is the PRESENTER (it did the work, a design doc, investigation, or analysis, and holds the real reasoning) and reconstructs the topic to a LISTENER, a spawned subagent pinned to a DIFFERENT-vendor model that has NOT seen the reasoning and interrogates it adversarially, hunting inconsistencies, gaps, unsupported magnitudes, and claims stated as fact but actually inferred. Goal, artifacts that are complete, gaps that are honest, findings that are truthful. Triggers include "rubber duck this", "rubber duck review", "explain this to a skeptic", "reconcile my findings", "is my doc honest".
Rubber Duck Review
Overview
The classic debugging trope, aimed at hallucination. The main session is the Presenter, it did the work and holds the real reasoning (including the hedges it never wrote down). It reconstructs the topic out loud, in its own words, to a Listener: a spawned subagent pinned to a different-vendor model that has NOT seen the reasoning and interrogates it adversarially. Two forces do the work:
- The Feynman effect, reconstructing the argument from scratch surfaces the Presenter's own hand-waving before the Listener even replies.
- An unstaked skeptic, the Listener has no investment in the conclusion, so it catches what the artifact's own self-review rationalized past.
This is a dialogue, not a static-artifact review: the Presenter reconstructs the argument live, and the divergence between the reconstruction and the written artifact is itself a finding.
When to use / when NOT
Use before shipping a findings doc, design, RCA, or investigation you want to be honest, anything where a "confirmed" might be hiding an unproven leap, or a magnitude might be hand-waved. Do NOT use for simple lookups or routine turns: it costs one+ extra model run per round and real Presenter effort. A deliberate, occasional move.
