The model that wouldn't be forced
My structured output worked by forcing Claude to call a tool that didn't exist. Opus 5.5 stopped playing along, and the fix was to stop doing the trick.
Tags
105 pages
Go
The commit message was the whole instrument panel
I asked an agent to diagnose a failing pipeline and leave it broken. It fixed it, then wrote a commit message about something it hadn't done, which bothered me a lot more than the fix.
A comment is for a different reader
The tool had comments on under ten per cent of its lines and ninety pages of documentation built into the binary, and I'd argue that was one decision made for two very different readers.
Every suppression is a claim
Forty-two comments telling a checker to stand down. I went through them one at a time to see what each was actually asserting, and some of them were wrong.
Read the report without ever holding the key
Publish an OpenPGP certificate a researcher can encrypt to, and read what they send, without a private key ever existing in your process.
Eighty-one megabytes of hello world
How big is a tool that does nothing? For go-tool-base it was 81 MB, and a comment in the code said it shouldn't have been.
My reason for hand-rolling it was wrong
I wrote my own OpenPGP packet code, then deleted the lot to prove I needed it. The reason I'd given turned out to be nonsense, and there was a better one waiting for me.
The library that appointed itself signal handler
A commit landed a month before I noticed it. By the time somebody told me, two of my own tools had already written workarounds for it without mentioning it.
A good signature in ten minutes
Ten minutes, a throwaway directory and no cloud account, to find out what a release signature actually is before you decide whether you want one in your pipeline.
The release that OOM-killed its own compiler
A wedding-photo culler ran out of memory compiling an LLM SDK it has never once called. The release had been running for thirty-eight minutes at that point.
