Set Core

set-core

An experiment in structured AI development — specs, quality gates, and parallel agents applied to Claude Code.

-lightgrey.svg)

### ⚠ This document is behind the code — last substantially revised 2026-04-27
Several capabilities have shipped since and are not described anywhere below:

Work on several of these is still in flight, and the documentation will be
refreshed once they land together rather than in pieces. Until then, treat
this README as accurate about what it describes and silent about the rest —
openspec/changes/ and openspec/specs/ are current.
This notice is here rather than absent on purpose. Stale documentation that
says nothing about being stale is the more expensive of the two states: it
reads as a complete picture, and a reader cannot tell the difference.

I use Claude Code every day. It's great for writing code, but shipping software — coordinating parallel agents, testing, merging, recovering from failures — needs more structure. set-core is what I built around it: give it a markdown spec, it decomposes into independent changes, dispatches parallel agents in git worktrees, runs quality gates on each, and merges the results.

Every change goes through OpenSpec — a structured workflow (proposal → design → spec → tasks → code → verify) that gives agents contracts to work against instead of prompts to interpret. Quality gates are deterministic: exit codes, not LLM judgment.

Built with set-core, using set-core. This project was developed using its own orchestration pipeline.
This is an early alpha — a working experiment, not a finished product. It works well enough that I use it daily, but expect rough edges. The sentinel auto-recovers from most of them. See docs/release/alpha-release.md for the current state, known issues, and what's not yet…