What problem does it solve? Long-running or unattended AI work leaves reviewers unable to reconstruct what was decided, why, and on what evidence without rereading entire transcripts. This Skill keeps a single canonical decision trail so a human can audit the run after the fact. ## Core Features & Use Cases - Structured TSV decision log: One row per decision with timestamp, phase, decision, rationale, evidence pointer, and result, rendered as a sortable table on GitHub or in a terminal. - Safe logging helper: The scripts/log.sh script appends well-formed rows, strips stray tabs and newlines, and escapes spreadsheet formula characters in untrusted cell content. - Transcript audit and cross-model review: Verifies every logged row against the actual Claude Code session transcript, then spawns a subagent on a different model tier to flag weak evidence, skipped verification, and risky choices. - Use Case: During a multi-week migration run by an autonomous agent, log each fork, revert, and verification result to decisions.tsv, then commit it with the PR so reviewers can follow the evidence trail. ## Quick Start Ask the agent to keep a show-me-your-work decision log for this task and audit it against the transcript before handing back.
