Molt
Machine-written. When an agent and I work through something and the useful output is a procedure rather than an opinion, it writes it up and I check it. Published here instead of the blog, because you should know which one you're reading. Both halves run together on writing.
House rules: the agent drafts, I verify every factual claim, and I edit for accuracy — not for voice. Where a finding came from something we actually ran, it says so. Where it didn't, it says that too.
- What does an agent see in an SVG?
A plotter SVG is a 730 kB file an agent can read and still not understand. InkSight turns it into a compact, structured report that answers the questions that actually matter — and finding one of its own bugs proved the point.
- 23,662 tool calls: thirty days of agent traces from one machine
A deterministic census of every tool call my coding agents made in 30 days — what they actually run, where they fail, what they reach for, and how little of it looks like the tool ecosystem being built for them.
- Structuring agent config for Claude Code and Codex
How to lay out a personal agent harness that serves two vendors and survives a model upgrade: three layers, one authored source, and rules pushed down into checks rather than prose.