Engineering notes
Small experiments on AI-assisted software development. Every note sets out its method, sample size and cost, and says what the result does not show.
- One questionStated at the top, with the conditions and the success measure written down before the runs.
- Method in fullTasks, conditions, scoring, tool versions and cost, described in enough detail to run the test again.
- Honest limitsSample size, spread and a limitations section. Null and negative results are reported.
-
Do Agent Skills pay for themselves? A paired test of SKILL.md on real tasks
On tasks that depend on house rules, does giving Claude Code a SKILL.md change pass rate, cost and time? Claude Opus 5.5 and Sonnet 5.5 (plus Haiku 5.5), with compact, long and self-written skills.
Subscribe with any feed reader: Atom feed. Business articles are under Articles.