Agentic Project Workflow — Data skill for Claude Code

🛠️ Agentic Multi-Repo Project Workflow

Take a piece of work across many repos from a pile of context to reviewed, shipped MRs — driven by AI agents, **gated by you at every phase**. It's a set of `/pw-*` slash commands plus a handful of templates, portable across agent CLIs (Claude Code, KiloCode, …).

**The idea, plainly:** say you need to change something across several repos — an upgrade, a new field threaded through three services, whatever. Instead of doing the thinking yourself, you hand an agent the raw material and let it work in stages: first it explains *what* needs to change and *why* (you read that and correct it before anything else happens), then it turns your approved understanding into a concrete task list (you approve that too — the one hard gate), then it actually makes the changes, each in its own disposable copy of the repo so nothing collides — and nothing is pushed anywhere until you say so. You're never surprised by a change you didn't see coming, and you never wrote the boilerplate.

**Terms used below:** a **worktree** is an isolated, disposable checkout of a repo (so parallel tasks never collide with each other or your own checkout) · **DAG** = the task dependency graph (which tasks must finish before others can start) · **MR** ("merge request", GitLab) = the same thing as a GitHub **PR** · an **agent** is the AI process you're driving; a **sub-agent** is one it spawns to do a single task in isolation (full definitions: [docs/EXECUTION.md](./docs/EXECUTION.md)).

 context  →  analyze  →  break down  →  execute  →  ship  →  close
  drop      what & why    PLAN + tasks  worktrees   MRs    learn +
  inputs    (you sign off) (you sign off) commit+verify    tear down

Every project lives in its own directory under `$PW_PROJECTS//` (wherever you cloned this bundle — see [ONBOARDING.md](./ONBOARDING.md), not a fixed folder name) — a copy of this bundle's [`template/`](./template). One phase writes, you review, the next phase star