Skills 2026-04-29 12:49 UTC
management-consulting
Committed by gcamilo. Update Category: Documentation Updates. Commit details: docs: refresh eval section with post-polish rerun (Apr 29). This update refines the documentation for easier onboarding.
CommitRaw commit message
docs: refresh eval section with post-polish rerun (Apr 29) - Add new "Latest: Post-polish (Apr 29 2026)" section with the skill-eval v2 rerun: skill_only avg 7.8 vs baseline 6.3 (+1.5), 3/3 wins - Per-tier deltas: T1 Process Discipline +2.2, T2 Output Quality +0.5, T3 Decision Quality +1.7 - Replace eval-results.png with the post-polish chart (Gemini-flash judge, Sonnet generation); fixes tier-label overlap from prior version - Demote pre-polish Opus blind A/B and earlier dual-judge numbers to "Pre-polish baseline (Apr 24 2026)" subsection — note that Opus run has not been re-executed against the polished skill The hard prompts (ambig-competitive, adv-framework-salad) score lower overall than pre-polish, but the skill is now correctly applying Quick Structure / refusing framework salad — Decision Quality on the framework-salad refusal scored 10.0. The rubric still rewards heavy structure; the post-polish skill trades structure for responsiveness when the prompt warrants it.