Quickstart

Before you start

• Claude Code for the enforced wall. Other agents can run the same procedure as instructions only; see other agents.
• Node 20+.
• For the rust-egui stack: cargo install cargo-nextest --locked.
• A lawfully licensed copy of the app you want rewritten, so you can take screenshots and recordings of it.

Run it

1. Install the plugin

   /plugin marketplace add remakecraft/remakecraft
   /plugin install remakecraft@remakecraft

2. Name the app

   /remakecraft:remake <app> --stack web

   Flag	What it does
   --stack rust-egui|tauri|web	The stack the rewrite is built in. Asked for if you leave it out.
   --builders 1-5	How many builders run at once, each in its own worktree.
   --cheap	Everything on Sonnet with one builder.
   --out <dir>	Where the run goes. Asked for if you leave it out; the default is a new folder per app.

3. Give it evidence

   Put your screenshots, recordings and the app's public docs into <app>.remake-room/evidence/. That folder is in the room: the observer and spec-writer read it, the builders never can. Add known open-source reimplementations of the app to blocked in remake.json so no agent can fetch, read or search them.

4. Watch the run

   Setup starts a live progress page on http://127.0.0.1:4317. It shows the phase, parity per layer with sealed results as counts, each feature, the agents and the wall log. It only binds to localhost and never shows a sealed test's name.

5. Approve the spec

   After the inventory, the spec-writer shows you the feature count, the must/should/could split and spec.md. Nothing is tested or built until you say yes, in chat or on the run page. Ask for changes and they go back to the spec-writer.

6. Approve the cost

   The pilot builds 10 must features, verifies them and reports tokens, time and pass rates per feature. From that it extrapolates the full run and asks whether to continue.

7. Read the score

   When the score stops rising, or you stop the run, you get the parity table from .remake/reports/parity.json and the wall log count. You can pause from the run page at any time; the run waits before the next build batch.

Crashed or hit a usage limit? Ask Claude to continue the run. It resumes from its progress files and git checkpoints, reading .remake/progress/checklist.md first and picking up at the first unticked line.

Models

Planners and reviewers run on Opus, builders on Sonnet, unless models in remake.json overrides them. The review runs through Codex CLI, or Claude if Codex is absent; models.codex picks the Codex model.

Other agents

The pipeline is agent-neutral, but the wall is not. Outside Claude Code the rules ship in AGENTS.md as instructions, and nothing enforces them.