Personal skill collection for coding agents
Corrects LLM bias toward human-team timelines. Forces decomposition into active agent time vs elapsed wall-clock time, grounded in METR and SWE-bench data.