Claude Code · Autonomous delivery

PRD to merge-ready PR

An AI engineering team that turns a plain-English feature request into committed, tested, reviewed code — without changing how product or leadership work.

3
Core AI agents
1
Human intent gate
0
Peer chat between agents
Full audit trail

The team

Four specialized agents orchestrated by /eng-team. Each reads from a shared scratchpad — no agent talks directly to another.

📋

Tech Lead

Reads the PRD and codebase, produces a precise Technical Spec — file paths, acceptance criteria, approach. Never writes code.

Spec only
⚙️

Engineer

Implements the spec, writes tests inline, runs the linter, and commits to a feature branch.

Build + test
🔍

Reviewer

Diffs against main for correctness, security, and performance. Approves with a PR description — or rejects with actionable fixes.

Diff review
🚀

DevOps

Deploys locally via Docker Compose. Separate /devops command after merge.

Deploy

How it works

Human provides the PRD. The autonomous pipeline runs spec → implementation → review, with bounded iteration loops when needed.

Human
PRD in
Autonomous
Tech Lead
Engineer
Reviewer
Merge-ready PR
Human merges

Bounded loops: Human ⇄ Tech Lead (PRD clarification) · Engineer ⇄ Tech Lead (spec gaps) · Engineer ⇄ Reviewer (reject/fix) · All state in .eng_team/task_*.json

Get started in four steps

Copy .claude/ into your project, bootstrap context, and run your first PRD.

Add eng-team to your repo

Copy the .claude/ directory to your project root. Requires Claude Code.

Generate CLAUDE.md

Run /eng-team-context ./ to analyze your repo and generate agent-optimized context automatically.

Run a PRD

/eng-team Add rate limiting to the cart API — max 100 req/min per user, Redis-backed, fail-open

Push & merge

Push the branch, open a PR, paste PR_DESCRIPTION.md. Deploy locally with /devops deploy local.

Built on a simple conviction

AI enters the engineering hierarchy from the bottom — where output is verifiable — and earns trust upward.

Verifiable first

Tests pass or they don't. Diffs make sense or they don't. Product judgment has no equivalent check — yet.

Narrow insertion

One slice: requirement → PR. Product, design, and deployment stay exactly as your team already owns them.

Trust compounds

Teams that watch AI ship reliable code earn the right to automate the next layer. Track record, not promises.