ai-workflow

AI-Assisted Development Workflow — Claude Code Plugin

Turns Claude Code into a small development team instead of one agent guessing alone — with a git-level gate that keeps the humans in charge of what actually ships.

What you get

• A team of specialists, not one generalist. planner, plan-critic, code-critic, and adversarial-qa each run in their own context window, built from one shared source of knowledge (the skills) — so each pass stays focused instead of one agent doing everything at once.
• /feature — one command, the whole lifecycle. Plan → critique → implement → test → code-review → QA → PR. You don't orchestrate the steps yourself; the workflow does, end to end.
• It stops and asks instead of guessing. When a critic raises a NEEDS_DECISION, or QA turns up something the plan didn't anticipate, the loop halts and hands you the decision — fix now, defer with a filed issue, or ignore. Nothing gets silently resolved on your behalf.
• A review gate enforced by git, not by an LLM's good behavior. A pre-push hook blocks pushing any commit that doesn't match a recorded, passing review — deterministic and outside the model's control. See Deterministic Enforcement in docs/AI-workflow.md.

Getting started

1. Install the plugin:

   /plugin marketplace add cunhaax/ai-workflow
   /plugin install ai-workflow@ai-workflow
   

   Developing the plugin itself? Point the marketplace at your local checkout instead, so you're testing your working tree, not the published default branch:

   /plugin marketplace add /path/to/this/repo
   /plugin install ai-workflow@ai-workflow
   

2. Scaffold and adapt it to your project:

   /init-workflow
   

   Fills in AGENTS.md (build/test/run commands, sensitive areas), points review and planning guidance at your existing docs (or seeds new ones), and validates the setup. Every write is proposed and confirmed — re-run it any time as a doctor.

3. Enable the review gate (once per clone):

   git config core.hooksPath githooks
   

4. Start a feature branch, then drive the work:

   /feature
   

   Start on a fresh branch you create yourself — the agents won't create or switch branches, and the workflow diffs against and opens a PR against your default branch.

Reviewing a small change yourself

Not every change needs the full /feature loop. For a small, manual session, you can tell the implementing agent up front that you'll do the review yourself — it will skip launching the code-critic sub-agent and, once you've looked over the diff and approved it, just run:

scripts/review-ok.sh          # records HEAD's SHA as reviewed

The pre-push hook only checks that the pushed commit matches whatever was last recorded — it doesn't care whether a sub-agent or a human did the looking. That's the same gate either way, just with you standing in for the critic.

Want the full picture?

Design rationale, the skill/sub-agent reference, and how the workflow executes step by step all live in docs/AI-workflow.md.

License

MIT — see LICENSE.