Writing
RSSEssays on this site
Twelve Checks That Could Not Fail
A reader pointed out that three characters in my release script were what made a broken check pass. He was right, and I measured it. Then I counted how many other checks in my own repositories had the same shape: twelve, in one week, all mine.
The Smoke Test That Could Not Fail
A release pipeline ended with a step that installed the published package on three operating systems and confirmed the version. It reported success against a version the registry was not serving, because the check looked for the version string anywhere in the output and the failure message quotes it back. Five more checks in the same repositories had the same shape.
Selected writing elsewhere
Humanizer has about 60,000 stars. On 12 public tasks my linter plugin scored 2.67 higher
ai-slop-linter against six other tools and no tool on 12 public rewriting tasks, graded blind, and the held-out set where the top tools tied.
(opens in a new tab)surviving-lines scored highest in a six-arm code-ownership benchmark
surviving-lines against five other ways of answering who wrote the code still in a repository, on public tasks and on a sealed set written by someone who never saw the tool, with the limits of each result.
(opens in a new tab)I made my coding agent explain every change to the customer. Here is the skill.
The product-engineer skill: every commit carries a plain-language section for the person paying for the work, and why that turned out to change what got built, not only how it was described.
(opens in a new tab)I open-sourced the part of a proactive AI assistant that decides when not to speak
The proactive-gate library: the suppressing half of a proactive assistant as a zero-dependency decision function, with the reasons it returns and the conformance fixtures two languages are held to.
(opens in a new tab)