agents-config
The configuration behind how I work with AI coding agents: one source rendered to four command-line tools, plus the skills, hooks and review gates that keep several agents building at once under control.
Private repository
01What it is
One repository holds everything I tell AI coding agents: instructions, skills, rules, subagent definitions and hook scripts. A bootstrap script written in Node.js renders that single source into the setup of four tools, Claude Code, Codex, Cursor and Antigravity. I edit a rule once and each tool reads the result. It has 60 skills and 308 commits since 25 May 2026.
02How a feature runs
I describe a feature to an orchestrating session. It settles the open decisions with me first, writes a spec with acceptance criteria and cuts it into slices. Each slice goes to a builder agent in its own git worktree, so several can build at once without touching each other’s files. On one recent feature that meant 18 decisions settled before any code, a spec with 25 acceptance criteria, 14 slices and up to five agents building at the same time.
03Who checks the work
A builder only builds: it compiles, runs the tests and writes a handoff note listing what to check in a browser. It never reviews its own work. A gate then checks the branch in a fresh context, with reviewer agents that did not write the code, a second review by a model from another vendor, Codex among them, and scripted checks. A check that could not run counts as a fail, never as a pass.
04Where I stay in the loop
Hook scripts block destructive git commands, keep a builder inside its slice and stop an agent from picking a more expensive model than its role allows. Merging and deploying stay with me, and pushing needs my explicit permission, given run by run. Every change to the workflow starts as a dated note in a research folder, with the question, the method and what could not be confirmed.
“If you haven’t seen it run, it’s not a working system.”
Deciding what to build and looking at the real screen stayed mine.
How it is built
- A feature request, planned and dispatched
- A finished slice, reviewed
- A checked branch, handed to me
- Configuration rendered to the tools
- Actions checked before they run
You
Decides and approves
Orchestrator
Plans the work
Builders
Write the code
Gate
Checks the result
Configuration
One source, four tools
Select a part of the diagram to see what it does. Each color is one flow through the system.