← Projects
airon icon

airon

Active

A shareable setup for Claude Code and Codex, installed once per machine — fifteen skills, the short rules that decide when each one fires, and the tests that measure whether a skill helps.

aiagentsclaude codecodexdeveloper toolingrustshelldotfiles

Working seriously with a coding agent means accumulating configuration, which is to say the instructions that it should always follow, and the skills that it can invoke, and the rules about when it ought to invoke them. That accumulation happens on whichever machine you happened to be using at the time, and then you sit down at a different machine, and none of it is there.

airon is that configuration kept as a repository and installed once per machine, so that every Claude Code session on it follows the same rules and has the same skills, and Codex shares them too. Changing how the agent works means editing a file in the repo.

What it gives an agent

There are fifteen skills, which are written workflows that the agent follows, and a short policy that says when each one is required, so that the agent uses them without being asked. tdd writes a failing test before it changes any code, kiss picks the simplest solution that works, investigate hands back a brief rather than answering from memory, and adversarial-review attacks a document before anybody acts on it. plan and orchestrate split a large change into streams and run each one as its own background session until its review merges. Others keep a knowledge base, track work that outlives a session, find what earlier sessions did, bring sessions back after a reboot, and run scheduled jobs.

Two of the skills test the others. skill-hardening gives a weaker model realistic practice jobs and nothing but the skill, and skill-shootout races a changed skill against its current version and against no skill at all, with a blind judge. Both are required before a change to a skill goes for review, and the follow-up post has what they found, including the three skills that were deleted because nothing ever used them.

Symlinks, not copies

The installer links the checkout into the .claude directory rather than copying it, so that improving a skill on one machine improves it in the repo, which means that it improves everywhere at once, rather than leaving 3 divergent copies behind. The settings file is generated rather than linked, because a couple of the paths genuinely differ from machine to machine, and the rules that load in every session are short stubs that point at their full text, because every session pays for whatever loads on each turn. A dry run shows every change the installer would make before it makes one, and the uninstaller never removes anything that airon didn't create.

Nothing personal in it

Everything about me lives in a private overlay repo instead, which is the memory that follows me between machines, and my own rules and settings, and the investigations worth keeping, and the list of jobs to run on a schedule. The installer layers it on top. airon's own rules and skills name nobody, and a test fails if the repo ever tracks a memory or names a home-directory path, which is what makes it shareable, although the repository itself is private.

It has a version number and a changelog now as well, so a session is told when there's a newer version out, or when it is running an older one than the one that's installed, and the agent is told to leave the update to me.

The generic cousin

airon is the generic counterpart of an internal setup that I maintain at work, and I split them on purpose rather than letting the split happen on its own. The half that is worth sharing is the plumbing and the skills, and the half that cannot travel is the integration with systems that only one employer has.

It began as configuration that sat underneath AgentRunner and TaskForge, and it has since taken on some of their jobs, because it runs scheduled jobs through launchd on its own, and its taskforge skill is moving the task record from the TaskForge CLI onto GitHub issues, which is built and still being verified.