Agentic Engineering Workflow
Turn your AI coding agent into a Director of Engineering for your team. Sprint cadence, multi-agent review triad, ADR discipline, and one-keystroke bug + idea capture.
A portable, opinionated workflow for software teams where an AI coding agent is part of the team. Distributed as a Claude Code plugin for one-command install — and the underlying skills + spec are agent-agnostic, so Codex, Cursor, Cline, and any other agent that reads SKILL.md-style instructions can use the same workflow with a manual install.
Drop it into any workspace and you get:
- 🎯 A setup-agentic-workflow skill that scaffolds
AGENTS.md, a CLAUDE.md shim, ADR structure, memory seed, tracker projects + labels — all adapted to your project's actual layout (greenfield, monorepo, or single-repo).
- 🐛
/defect — fire-and-forget bug capture from chat. Infers platform, build, repro steps, severity, root-cause hypothesis, and likely-culprit files. Files a tracker ticket with all of it.
- 💡
/idea — parking-lot idea capture. Infers cross-repo scope, drafts a mini-spec, files it for future sprint triage.
- 📚 A full workflow spec (sprint phases, triad review pattern, contract-lock discipline, working agreements, common pitfalls) you can read once and reference forever.
Tracker-agnostic (Linear, GitHub Issues, Jira, Notion). Stack-agnostic. Agent-agnostic. Works solo, scales to small teams.
Install — Claude Code (one command)
/plugin marketplace add Pixel-Perfect-Apps/agentic-engineering-workflow
/plugin install agentic-engineering-workflow@agentic-engineering-workflow
After install, ask your agent to "set up the agentic engineering workflow" in any workspace, or invoke /agentic-engineering-workflow:setup-agentic-workflow directly.
/defect and /idea are available globally after install — no project-by-project setup.
Install — Codex, Cursor, Cline, or other agents (manual)
These agents don't share Claude Code's plugin format, but the skills + spec are portable. Two options:
A. Per-project install (Codex .agents/skills/ convention):
# From your project root:
mkdir -p .agents/skills
git clone --depth 1 https://github.com/Pixel-Perfect-Apps/agentic-engineering-workflow /tmp/aew
cp -r /tmp/aew/plugins/agentic-engineering-workflow/skills/* .agents/skills/
rm -rf /tmp/aew
Then point your agent at .agents/skills/setup-agentic-workflow/SKILL.md to kick off the setup. The skill content is generic — it works for any agent that follows SKILL.md-style instructions.
B. Drop-in spec only (no skills, any agent):
Save just the spec into your workspace and paste the setup prompt to a fresh agent session:
curl -O https://raw.githubusercontent.com/Pixel-Perfect-Apps/agentic-engineering-workflow/main/plugins/agentic-engineering-workflow/skills/setup-agentic-workflow/references/team-process-spec.md
Then open the file, find the ⌜ PROMPT TO PASTE ⌟ block, and paste it to your agent. The agent reads the spec and executes the same five-phase setup. You miss the /defect and /idea slash commands but the rest works.
What the setup does
The setup is inspection-first and consent-based. It refuses to write to your tracker or files until it has shown you what it found and asked what you want.
The flow:
- Inspect, read-only. Walks the file tree, queries the tracker (projects, sprints, labels, built-in Triage availability), looks for existing docs locations, and — critically — detects whether another collaborator has already run the setup here.
- Asks one batched round of questions. Leads with either install-state scope (sync me locally / complete partial setup / update team config / re-install) if it detects existing setup, OR workspace-shape classification if fresh.
- Scaffolds files the user approved —
AGENTS.md (with embedded "Tracker destinations" + "Docs location" sections), CLAUDE.md shim, per-module stubs, optional ADR placeholder at the chosen docs location, local memory files.
- Writes to the tracker only what the user explicitly approved — never auto-creates "M0 sprints" or planning artifacts; uses existing destinations when they match; adds labels only with consent.
- Reports what was written, what was skipped, and the next prompt to run.
Concretely, you'll end up with (depending on what you approved):
