agent-memory-designerlisted

CLAUDE.md is loaded into every prompt for the whole session, so it is the highest-leverage and most easily-abused file in the repo: a sharp 40-line memory steers every turn, while a 400-line dump of prose gets skimmed, diluted, and quietly ignored. This skill explores the actual project, decides what earns a permanent slot in context, and writes a memory the model will actually follow. ## When to use this skill - You're onboarding an existing repo to Claude Code and there is no CLAUDE.md (or `/init` produced a generic one that restates the obvious). - The current CLAUDE.md is long, stale, or contradicts the code, and Claude keeps running the wrong test command or ignoring a stated rule. - You want to split repo facts (project CLAUDE.md) from personal cross-project preferences (user `~/.claude/CLAUDE.md`) and aren't sure what goes where. ## When NOT to use this skill - You want an automatically enforced rule (format on save, block edits to a path) — that's a hook, not memory; memory is advisory and the model can deviate from it. - You want a reusable procedure with its own tool scope, invoked on demand — that's a skill, not a fact that must sit in context every turn. ## Instructions 1. **Read the repo before writing a word.** Glob for the build manifest (`package.json`, `pyproject.toml`, `Cargo.toml`, `go.mod`, `Makefile`) and read its scripts/targets to get the *real* commands — never invent `npm test` if the script is `npm run test:unit`. Grep configs (`tsconfig`, `.es