I direct AI agents, and most of what they build is tooling for directing AI agents better. Almost none of the code in these repos was written by me. That is the point of them, not a disclaimer.

254 of the 256 content commits across the four repos below carry a Claude Code authorship marker. The two that don't are an MIT licence file and a bulk first import of files an AI had already generated — so the honest summary is not "mostly", it's essentially all of it.

Don't take my word for it — clone any of them and run:

git log --no-merges --format='%an|%(trailers:key=Co-authored-by,valueonly,separator=;)' | grep -ci claude

If anything the number understates it: a commit made under my own git identity can still be AI-written, and plenty are.

What's mine is the part git log doesn't record — choosing the problem, writing the standard down before the model starts, and binning what doesn't meet it. Whether that's worth much is a fair question, and these repos are exactly where you'd check, because the gates either go green or they don't.

So this profile isn't look what I built. It's look what this produces when someone holds it to a standard instead of shipping the first output. Judge it on the artifacts.

A short scoping interview, three mockups to pick from, then a production-grade Next.js 16 site — built through the review checkpoints a real studio would give you, with preview URLs you check on your own phone.

80 skills · 3 model-routed subagents · 7 screenshot-verified quality gates · a written taste constitution that every other skill defers to.

The showcase — ultraweb-site — was built by that pipeline from one prompt, with no human touch-ups. Every decision it made on the way is committed in design/*.md, so if the site were bad you could prove it from the repo: ultraweb-site.vercel.app

Long-horizon agent work fails in repeatable ways: declaring victory without running the check, git reset --hard over uncommitted work, grinding the same failing command, losing the original request across a compaction. Advice loses to momentum — so the load-bearing rules sit behind hooks that cannot be talked out of, and the checks that matter go to fresh-context agents that owe the work no loyalty.

python tools/demo.py runs the actual shipped hooks against planted failure modes and asserts each one blocks. CI runs it on every push.

gtheme — a GNOME theme manager for people who are nervous about breaking their desktop. Wallpaper, colours, icons, pointer, shell and add-ons in one window, explained in plain words, with an undo that always works. Pure-Python CLI + TUI, no telemetry.

NightCityMP — multiplayer for Cyberpunk 2077 (patch 2.31a). Host and join like Minecraft. A continuation of CyberpunkMP.