A delivery harness, not another chat window.
Ralph Starter packages the operating system around AI coding: specs, isolated worktrees, race and consensus modes, integrations, verification, and a clean path back into a real repository.
Raw model capability was not the bottleneck. The hard part was making parallel agents useful inside repositories where collisions, stale context, unverifiable claims, and unbounded token use can erase the speed advantage.
- 01
Specs define outcome, constraints, acceptance checks, and ownership before work starts.
- 02
Each worker runs in an isolated git worktree so parallel attempts remain comparable and recoverable.
- 03
Race, consensus, and pipeline modes match the collaboration pattern to the problem instead of treating every task alike.
- 04
Repository checks and visual evidence decide what lands; the model does not grade its own output.
The result is a reusable operating layer for AI development: teams can increase parallelism without giving up repository discipline, reviewability, or the ability to reproduce how a change was produced.