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

Essentially every commit that touches the code in the repos below was made with Claude Code, and the git history says so: Claude is named as author or co-author on nearly all of them. Clone any repo and read the log.

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. These repos are where you'd check whether that's worth anything, because their 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. Judge it on the artifacts.

A Claude Code plugin that builds production Next.js and Tailwind sites through a guided design session. Describe the site in a sentence, answer a few scoping questions (never about colours or fonts), pick from a handful of deliberately different mockups, and the pipeline builds, tests and ships the rest. By default it pauses at mockup approval, first-page review and final acceptance. When Vercel is connected, the first-page review and final acceptance come with a preview URL you can open on your own phone.

Nothing ships on the model's say-so. The gates run against a production build in a real browser: compile and typecheck, responsive screenshots, an adversarial design critique, accessibility, performance, an anti-slop sweep for purple AI gradients and "Empower your workflow" copy, and a hunt for placeholders and dead links. A failure never ships; a gate that could not run holds the release until you accept each named risk.

The taste skill is a written design constitution every other skill defers to. Every phase writes its decisions to design/*.md, so every decision is on record.

The showcase, ultraweb-site, is what that pipeline produced from one prompt with no human touch-ups. The repo holds the source and the full paper trail, and the site is live at ultraweb-site.vercel.app. If anything had been fixed by hand, the history would show it.

Long-horizon agent work fails in repeatable ways: declaring victory without running the check, wiping uncommitted work with git reset --hard, grinding the same failing command, losing the original request across a compaction, handing work to a "verifier" that can quietly edit it. Advice alone 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 are read-only by hook enforcement, not by promise.

One of those hooks, the claim-audit gate, reads the transcript for evidence that a check ran and passed after the last edit. A per-session ledger records whether those hooks actually fired, so "the floor is armed" is measured rather than assumed. python tools/demo.py runs the shipped hooks against planted failure modes in a throwaway sandbox and asserts each one behaves; CI runs it on every push.

gtheme — a safe way to change how your GNOME desktop looks, for people new to Linux. Wallpaper, colours, icons, pointer, fonts, top bar and add-ons in one window, every switch explained in plain words. Everything is saved first, changes are all-or-nothing, Ctrl+Z undoes from anywhere, and a "before gtheme" snapshot is kept forever. Its whole-desktop presets, called Looks, are settings rather than code, so they cannot run programs. No account, no server, no telemetry.

NightCityMP — multiplayer for Cyberpunk 2077. Host and join like Minecraft. A continuation of CyberpunkMP.