← Luke Hornof

Projects

A growing wall of builds. Each one records how it was made — the stack, the agents and models, the tests, the cost — because how a thing gets built is half the story.

Everything here came out of the same loop: I write a spec and gate it, agents build and check each other's work, I review every pull request, and nothing final until the deployed result is verified. The poster is that machine on one page.

Hand-drawn poster of the agentic programming workflow: a cast of the human director/owner and the agents — Cowork the read-only verifier, Claude Code the programmer, background workers, and a scheduled-task babysitter; the eight-stage build loop drawn as a winding road with human approval gates at the PRD and engineering doc; a panel of the loop's hard-won rules; and a shared-brain table listing the artifact files and their owners.
The method (generated by NotebookLM from my process guide, modified by Claude agents).
The agentic programming workflow poster, shown at full size.

hornof.org

2026-07

This site. Rebuilt from a Dotster one-pager into a modern static site by an agentic engineering loop — each feature specced, tested, and shipped as its own reviewed pull request.

Built with Astro · static HTML · content collections · one style.css · Cloudflare Pages · Claude Code · 120+ tests · first-pass green: mostly — human review caught the restTests written before code; every feature landed as its own reviewed PR, none straight to main.

eclipse-sim

2026-07

A solar-system orrery computed from Newton's laws — no ephemeris lookups — that predicts real solar and lunar eclipses and shows its miss margin against the NASA catalog on every row. Built overnight from a one-page brief.

Built with ES modules · Three.js (CDN) · Web Worker · no bundler, no build step · Claude Code (/goal loop, unattended overnight) · 135 tests · first-pass green: overnight suite green — then hand-testing found 4 bugs the suite missed; see the writeup47 minutes from brief to working predictor. The debugging afterward became the testing playbook (property-based + mutation tests, 94% mutation score).

eclipse-sim Open full screen ↗