I write on Medium about what the logs and benchmarks show. Newest first.
in Turkish
How a 270M Turkish PII masking model went from 0.740 to 0.882 on a public 1,000-row benchmark with 24,000 targeted examples, and the two slices where it got worse. Related: the model page.
Read on Medium
A crawl, walk, run guide for every team in a large company, not only the technical ones. One team cut its daily incoming requests from 100 to 35 without buying an AI tool, by charting its own data first.
Read on Medium
Claude Code writes down which skill was active on every tool call. Across 128 sessions there were 376 activations. Six skills did all the work, and 1,447 lines of instructions were never opened by any agent. Related: skillbench.
Read on Medium
Instruction files are read before every question you ask, and nobody measures them. In twelve projects the median was about 6,000 tokens per request, and one rule turned out to be working because it never fired. Related: gardener.
Read on Medium
Claude Code writes every session to your disk as JSONL. Mine showed 35 tool calls per prompt, 49 denials and other things that did not flatter me. Related: devpersona.
Read on Medium
New articles, new plugins and news about the hosted version, by email. At most one message a month, and you can unsubscribe at any time.
Or follow on Medium.