About
I’m JJ. Peru ↔ New Jersey, 16 years building enterprise software, currently leading engineering teams.
The agent layer is commoditized. Claude, Codex, Cursor and Gemini are racing each other, and whichever one is ahead this month will not be ahead next month. My time goes into the layer above: tooling and institutional knowledge that compound across model swaps, the kind you don’t have to rebuild when you switch harnesses.
Stop building another Claude. Build the harness around the one you already have.
What I write about
Using AI to build software, tested rather than theorized. Every post comes out of something I actually shipped: a tool I built and use daily, a workflow that survived contact with real work, or a number I measured on my own machine. When a claim has a source, it links to the source. When it’s my opinion, it says so.
The recurring threads are second-brain workflows for agents, the economics of AI-assisted engineering, and the harness layer that sits between a model and useful work.
What I build
Everything is open source and lives on GitHub, which is the canonical list rather than a copy of it:
- voice-tunnel — talk to your coding agent from your phone, all local, no GPU
- ccburn — real-time burn-up charts for Claude Code usage limits
- hwinfo-tui — plot HWInfo64 sensors live in the terminal
- atref — Claude Code’s
@file picker, in any text field - agent-mail-cli — a mailbox coding agents can learn in one command
- claudefana and claudefana-enterprise — dashboards for Claude Code telemetry, personal and org-wide
- claude-code-tips — 14 habits for shipping more and spending less
Subscribe
A new post goes out every Tuesday. The RSS feed has everything, and every post page carries a subscribe box if you’d rather get it by email.