workflow-distillerlisted

# Workflow Distiller The meta-skill: your transcript history already contains the skills you should have written. Every workflow you've re-explained to Claude three times, every shell ritual you've re-typed, every correction you've made twice — those are skills waiting to be distilled. This skill scans your history in aggregate, clusters the recurring patterns, scores them, and drafts the SKILL.md files for the winners. The skill has three phases, and they never blur: 1. **Scan** — a small auditable Python script emits aggregates only. 2. **Judge** — Claude clusters the aggregates into named workflow candidates and scores them. 3. **Draft** — the user picks; one SKILL.md per pick, written from the pattern, never from the text. ## Privacy Contract (present this BEFORE anything runs) This section is read to the user verbatim or paraphrased faithfully, before any scanning happens. It is the first thing that happens when this skill fires — before any tool call, before any file read. **What is read:** - `~/.claude/projects/*/*.jsonl` — your local Claude Code session transcripts - Project directory names (e.g. `-Users-you-Desktop-myapp`) - Nothing else. No other files, no network calls, no telemetry, no uploads. **What the scanner emits — aggregates only:** - Counts: sessions per project, messages per project, malformed lines skipped - Recurring command n-grams: tool-name 3-grams and, for Bash, the first token of each command (the binary — `git`, `npm`, `ffmpeg` — never th