I write hooks and verification loops for coding agents.
Claude Code, Codex and Kimi each run the same task. I publish the runs that fail.
The tooling I build sits around the model rather than inside it. A hook fires before your prompt reaches the model and injects what you already decided. A grader with a fresh context reads the diff without ever seeing the conversation that produced it.
fable-mode · Python Run Claude Fable 5 behavior on Opus 4.8 inside Claude Code. A native distillation of how Fable operates as an agent, plus a measured execution playbook, verification hooks, and design and test skills. One clone, one install.
claude-recall · Python A memory-recall hook that injects your past decisions into the prompt. I built the vector database version first, measured it, and tuned lexical search beat it on my vault. So this is one file, standard library only, no MCP and no API calls.
homelab-voice-assistant · TypeScript You speak, the right thing happens in about a second, and the screen shows what was heard and which tier decided. whisper.cpp for speech, a tool-calling model on Ollama for the brain, Kokoro for the voice. No Anthropic API key anywhere in it.
tor-vpn-macos · Swift, Shell Tor in Docker with a native menu-bar toggle. No kernel extension, no sudo, no routing-table surgery. Benchmarks public obfs4 bridges by measured throughput when your network blocks Tor outright.
Every repo here runs against my own work daily before it reaches you. When a run fails I write down what broke, then ship the fix as a rule the harness enforces. A note I promise to remember is not a fix.