Shreyp087 Retrospective loop
Use actual event artifacts, timings, evaluation results, judge feedback, and team notes. Separate observations, participant interpretation, and missing evidence. Winning or losing one event does not identify a causal explanation by itself.
Compare estimates with actual time for discovery, setup, implementation, integration, debugging, and submission. Inspect the largest delays and failures. Track time to first working slice, reproducible demo success across rehearsals, evaluation outcomes, and manual rework when those were actually measured.
Identify one to three changes supported by the evidence: a clearer trigger, a reusable fixture, a simpler interface, a missing verification, or removal of an unnecessary step. Prefer a narrow correction over another global rule. Record tradeoffs and the situation where the change applies.
When changing a skill, use the available skill-creator guidance, preserve the user's intended workflow, and run the relevant validators. For a meaningful behavioral change, replay a realistic task in a temporary workspace and compare the observed result. Never run a live external mutation solely to test a skill.
Deliver what happened, what changed, why the evidence supports it, and what remains uncertain. Maintain a dated decision log. Retire unused skills or move rarely needed detail into references when discovery becomes cluttered; more installed instructions are not a productivity metric.