// projects

Stable

Live

A small stable of agents and skills for Claude Code, curated by hand. Every piece has to earn its place in the context window, and Garry, the dispatcher, never merges on his own.

*The catalogue, as install.sh --list prints it: the three agents first, then the skills*

*Garry, in his own words. He carries the brief and the diff, and he never merges without being told to*

*The guardrail. A plain question gets a plain answer, with no planning ceremony*

1 / 3

// summary

Most agent tooling ships in bulk. You install two hundred agents, and the model spends half its attention working out which one you meant. Stable goes the other way. It has three agents and fifty-seven skills, all vendored so I can edit them, and one Bash script installs any or all of them.

The interesting parts are the guardrails. I rewrote a popular workflow skill so it stops launching a full planning ceremony when someone simply asks a question. I also built Garry, a dispatcher that hands a task to one AI model in an isolated git worktree, then asks a different model to review the diff without touching it. Garry never merges on his own. In testing, the reviewer was told to go ahead and fix things itself, and it declined, which is the most obedient thing any reviewer has ever done for me.

There is also a narrow specialist for signal-processing maths, built for my audio side projects. It is scoped to catch filters that run fine and are quietly wrong, and it is under strict orders to have no opinions about tone.

// tech stack

AI / ML

Claude Code Claude Opus OpenCode Codex CLI Pi

Language

Bash Markdown JSON

Tooling

Herdr Git worktrees jq