kept-promises

What problem does it solve? When an AI agent claims it implemented everything you asked for, there is no way to verify that claim without manually digging through git history. This Skill produces a count, not a promise: it reads the transcript of every instruction you gave and checks git for evidence that work actually landed afterward, flagging anything that was dropped, thinly done, or claimed with a commit sha that does not exist. ## Core Features & Use Cases - Instruction extraction: Parses Claude Code session transcripts and classifies owner turns as instructions versus questions, acknowledgements, navigation, or pasted exercises. - Four verdicts per instruction: KEPT (a later commit matches strongly), WEAK (something matches thinly and needs human eyes), NO TRACE (nothing in history matches), and PHANTOM (a commit sha was quoted in a reply but does not resolve in git). - Sha-linking over keyword guessing: Commit shas quoted in assistant replies are verified directly with git cat-file, so a fabricated commit reference becomes the loudest possible finding. - Use Case: At a project milestone, run the full audit to get a count of kept, weak, and untraced instructions across all session transcripts, then review the WEAK and NO TRACE lists before trusting any completion claim. ## Quick Start Ask the agent to run the kept-promises audit across all transcripts and show which of my past instructions have no trace in git.