Labs
When I have a question about working with AI, I’d rather test it than guess. Each lab starts with a concrete question, puts it to the test and ends with a verdict that fits in one line. The method is on the page and the code is in a public repository, in case you want to check it or run it yourself.
3 labs
confirmed refuted inconclusive
Does a design skill beat using no skill at all?
The gap between skills was smaller than the gap between two runs of the same skill.
Do subagents cut context use on a legacy migration?
41% fewer tokens in the main thread, with the same test pass rate.
Is hybrid retrieval worth it below 10,000 documents?
BM25 alone matched hybrid on nDCG@10 at a third of the latency.