I went looking for token waste in my AI harness traces. Instead I found behavioral failures: agents that fear complexity, write stubs and declare victory, and silently delete code they don’t understand. I built four mechanisms to stop them.
Stanford’s Meta-Harness paper argues that optimizing the code around the model matters more than optimizing the model itself. I applied it to my Claude Code setup and got the first real numbers: 3,723 tokens always in context, with no evidence they’re the right ones.
Everyone says AI makes developers faster. My experience is the opposite: I arrive exhausted not because I wrote more code, but because I evaluated more decisions. The effort didn’t shrink. It transformed.
Three layers of knowledge integration between Obsidian and Claude Code: from write-only vault to a feedback loop that promotes patterns into skills.
Structured requirements, deviation rules, and the art of selective theft: five concepts borrowed from GSD, twenty rejected.
From token optimization to an orchestrated development workflow with Claude Code: teaching an AI your patterns is only part of the problem.
A personal project that started as “can I automate an RSS digest?” and became a full-stack monitoring platform with narrative memory, semantic search, and AI-generated editorial commentary.
Another chapter in our automation journey at Wishew, where we bridge the gap between project management and development workflows, and discover what happens when your AI pair programmer actually knows your patterns.
The best way to work with AI isn’t to ask it to write code: it’s to teach it your patterns first.