claude-put-in-work

A framework for autonomous, multi-session app building with Claude Code. Describe what you want built, run build.sh, and let Claude build it end-to-end — overnight if you want. In v3, a Strands-powered reviewer agent watches the build from Discord: it reviews every phase, runs the tests, screenshots the app, and holds the final phase to an approval gate — while you steer the build from your phone.

The diagram is drawn from assets/build-loop.html — edit the HTML and re-screenshot to keep it truthful.

How It Works

 1. PLAN                2. BUILD                        3. REVIEW

 Your idea/plan          ./build.sh ──────┬────────> ./reviewer.sh (auto)
       │                        │         │                  │
       v                        v         │                  v
/autonomous-build     ┌───────────────────┐       ┌──────────────────────┐
       │              │ Claude Code loop   │       │ Reviewer agent        │
       v              │ session 1 -> 2 ... │       │ (Discord + Strands)   │
┌─────────────┐       └────────┬──────────┘       └──────────┬───────────┘
│  PROMPT.md   │               │                              │
│  build.sh   │        after each phase:              reviews the phase:
│  reviewer/  │        appends "phase: N"             checklist + tests +
└─────────────┘        to .review/queue ────────────> screenshots, then
                               │                      posts a verdict
                               │                              │
                               v                              v
                     ┌──────────────────┐          ┌────────────────────┐
                     │ BUILD_PROGRESS.md │<─────────│ NEW_FEEDBACK.md     │
                     │ (session handoff) │ feedback │ .review/verdicts.md │
                     └──────────────────┘          └────────────────────┘
                     "ALL PHASES COMPLETE"          final phase requires
                      only after APPROVAL <───────── an APPROVED verdict

1. Plan — Describe your idea and run /autonomous-build. The skill generates a PROMPT.md spec, the build.sh runner, and the reviewer agent in your target project. Review the spec before running — it's your last checkpoint before autonomous execution.
2. Build — Run ./build.sh. Claude Code runs in a loop, reading BUILD_PROGRESS.md each session to pick up where it left off. No human intervention needed until it's done.
3. Review — If Discord is configured, build.sh also starts the reviewer. Every completed phase gets reviewed with a verdict posted to your channel, you can chat with the reviewer about the build anytime, and the final phase can't complete without approval.

Two ideas make this work: BUILD_PROGRESS.md gives Claude continuity across sessions, and the .review/ file contract lets a second agent supervise the first without either one blocking.

The loop from your seat

In practice this is two terminals: a strong planner in one, the builder loop in the other, and you deciding when it's done.