Writing

Notes from the desk.

Longer pieces about the things on this beach and how they came to be. Written slowly, in the order they bothered me.

A number instead of an opinion

the eval harness behind foresound's tagging

How we built an eval harness for Foresound's music tagging: a framework where models and prompts are variables, gold sets labelled by about 800 agents across Claude Code and Cursor, a cleaner data model, and a model that matches production on artists and genres for about 62% less per post.

sep '26 · velais for foresound

A blind test for agent skills

what skill arena measures, what it can't, and why there is a cat

Skill Arena runs the same prompt with and without an agent skill and makes you vote before it tells you which was which. Notes on building it, and on the cat.

sep '26 · built on glaze

← back to the beach