What problem does it solve? AI agents are poor judges of their own completeness, so sessions often end with unverified assumptions, skipped checks, or dropped threads that silently invalidate the work. This Skill forces a structured confidence audit and blind-spot check at the cheapest possible moment: right before the user walks away. ## Core Features & Use Cases - Two-question reflection protocol: Sequentially asks "What am I least confident about?" (3-7 concrete, file-level items) and "What's the biggest thing I'm missing?" (1-3 broad patterns), never combined into one message. - Inline or sub-agent modes: Runs inline for short sessions, or spawns a sub-agent with a concrete session summary for deep multi-file sessions, with guidance on cheap model selection. - Gap persistence: Converts verified-but-deferred findings into known-issue documents via the context-file skill, with honest severity ratings, instead of leaving them as chat mentions. - Use Case: After a multi-file refactor touching retry logic and DynamoDB writes, the user says "we're done." The Skill triggers a reflection that surfaces an unverified GSI assumption, which is then investigated or filed as a known-issue before the session closes. ## Quick Start When the user signals the session is wrapping up, initiate the reflection by asking what you are least confident about, wait for the response, then ask the blind-spot question.
