Lab
Experiments, logged
Each entry is a testable question about working with agents, the setup I ran it in, the evidence produced, and what I now believe. Written as a log rather than an article, so a failed run is still a result.
EXP-001EXP-001 - Porting 151 posts off a Next.js blog with an agentAn agent can migrate a 151-post legacy blog to a static site with full URL preservation and zero content loss, if the migration is decomposed into crawl → normalise → verify instead of attempted as one task.running16 NOV 2026EXP-002EXP-002 - Does a written plan file actually improve success?Committing a plan file before implementation raises first-pass success on multi-file refactors, and lowers the human review time per merged change - even when the plan is later wrong.idea10 NOV 2026
How the lab works
- One question per entry. If it can't be phrased as "does X change Y", it becomes an essay instead.
- Setup is written down before the run - model, harness, repo size, task, budget.
- Raw artifacts are kept - prompts, transcripts, diff stats, cost. Linked, not summarised.
- Negative results get published. A run that proves nothing is still evidence about the method.
- Status is honest: idea queued, running in progress, pass / fail concluded.