Instead of hand-fixing an AI draft’s awkward sentences every time, I built a correction loop where one flag becomes a permanent rule. The Claude Code skill structure, how real rules were born, layering a community skill on top, and a seed file you can start with in ten minutes.
One paragraph of Tolkien, a million tokens, two hours of autonomous execution, a 3D film. Not a demo review: how the unit of delegation is changing, and where a builder’s work moves next.
I build an agent memory platform, and this week we reversed its direction. What store-everything memory breaks, and how to route state, facts, and principles to the tracker, the code, and memory.
A CLAUDE.md does not get better as it grows. Why piling up rules degrades your agent, what belongs in the file, and where the rest should live: skills, hooks, and memory.
I built Monet to fix the agent that forgets yesterday’s conventions. How an agent that writes its own memory, reads what mattered, and grows with use actually works.
Two developers, same tool, wildly different setups. How deep coding agent customization goes — from bare defaults to full pipelines — and what actually pays off.
What today’s LLM benchmarks actually measure, what they miss, and why we need behavioral benchmarks instead of more knowledge tests.
Two requests with near-identical input tokens, a 46x cost gap. How prompt caching really works, and why agents should be designed around structured state.
Coding agents re-read the same files and re-ask the same questions every day. What memory actually does for an agent, how it differs from docs, and when it pays off.
Bigger context windows didn’t give better results. Notes on the context-management problems that keep recurring in real coding agent sessions.