onboardlisted
Install: claude install-skill yoelgal/agent-tools
# Onboard a repo into better-dev
Bring better-dev into this project - greenfield or an existing codebase - by detecting what's
already here, adapting to it, wiring memory, and leaving a discovery block so every later session
knows the practices are available. One job: **get the repo wired, without imposing.**
better-dev installs in two layers, and onboard only touches the second. The **tool** (the skills +
`bd-*` scripts + hooks) is installed once per machine, globally, into your host's native skills dir
(Claude Code: `~/.claude/skills/<skill>`; Codex: `~/.codex/skills/<skill>`, one symlink per skill), never
vendored per repo. This repo carries only **data**: `.better-dev/` with rules, overrides, and
learnings - tracked and shared on a team adoption, local-only on a solo one (Phase 2 asks which) -
plus a per-machine `.better-dev/bin` symlink back to the global install so the portable
path `.better-dev/bin/bd-mem` resolves here.
## Agent contract
Run the phases in order. If `$ARGUMENTS` names a phase, jump straight to it. At each phase:
**detect → report tersely → confirm → act.** Skip anything already done - this skill is idempotent
and safe to re-run; re-running only fills gaps.
Three rules carry the whole skill:
- **Detection is a premise, not a fact.** Verify it at `file:line` before you build on it. A branch
named `staging` in `CLAUDE.md` prose isn't a `staging` branch until `git` shows it. Report what you
actually observed, with where.
- **Never guess a command