I build small, sometimes useful tools that make AI assistants useful at home and work: prompt libraries, MCP servers, and Claude skills that encode how a team really operates.
Most of what I write turns brand standards, policies, and tribal knowledge into things a model can follow. The rest is me automating whatever is annoying me.
- Lapidary: Create your best prompts once, improve them with tests, turn them into assistants, and use them from Claude, ChatGPT, Cursor and Open WebUI, on a free local model or your own key. Self-hosted, with a Windows 11 console.
- Claude skills for teams: packaging brand, compliance, and workflow knowledge so the assistant gets it right the first time instead of the fifth.
- Debloating a 2018 Samsung NU8000: stripping the stock apps and swapping in a cleaner launcher, with every removal logged so it can be undone.
| Project | What it is | Built with |
|---|---|---|
| Lapidary | Self-hosted prompt console for Windows 11 | Python, FastAPI, FastMCP, Next.js, TypeScript, Electron, SQLite or Supabase, Ollama, PowerShell 7 |
- Docs first. Every repo ships with
CONTEXT,ARCHITECTURE,DECISIONS, andROADMAPfiles so agents and humans start from the same page. - Agent-friendly by default. If a tool can't be driven from an MCP client, it isn't finished.
- Boring infrastructure on purpose. Postgres, Docker, and a test script before anything clever.
- Reversible changes. Log what you removed, keep the backup, and leave a way back.
