skillet-createlisted

# Skillet create (`/skillet create`) Most people cannot answer "what skill do you want to write?" They can answer "yes, that one" when you show them something they already do. So this playbook never opens with a blank page. It reads the work that already happened, proposes candidates with evidence, and carries the one they pick all the way to a live skill page. Six phases, in order. Do not skip, do not merge, do not run more than one skill per invocation. ## 0. Say what you are about to read One line, before you read anything. Not a privacy essay, a scope statement: ``` Reading this session, this project's recent transcripts, and its conventions files to find what you keep doing. Stays on your machine. ``` Then read. Everything in phase 1 is local. Nothing is sent anywhere until the user approves an upload in phase 4. **Never widen the scope on your own.** Other projects, other repos, and shell history are out unless the user names them. If the current project yields nothing, ask before reaching further. ## 1. Mine the evidence Work the ladder top-down. Stop when you have enough for three honest candidates. 1. **This session.** Already in your context and free. What did the user correct you on? What did they explain that they clearly explain often? 2. **This project's transcripts.** Claude Code keeps them at `~/.claude/projects/<slugified-cwd>/*.jsonl`; Codex at `~/.codex/sessions/`. Grep for repeated instruction shapes rather than reading whole files: "