mirror of
https://github.com/pbakaus/impeccable.git
synced 2026-09-14 15:16:35 +03:00
Paul BakausandClaude Opus 4.7
a35c928183
Bisection across 5 historical skill commits on Gemini 3.5 flash fast lane n=3 found that0cf2debdwas the peak quality state. The regression between0cf2debdand HEAD came from simplifying the long anti-cream paragraph into a one-liner. Restoring the paragraph (with em-dashes replaced by parens to satisfy prose lint) recovers ~0.22pt average on Gemini vs HEAD, with the largest gains on: - 09-luxury-hotel: +0.50 (restores editorial drama in photo-led briefs) - 10-food-magazine: +0.67 - 03-italian-restaurant: +0.51 The paragraph's load-bearing parts are the (a)(b)(c) alternatives that give the model actionable replacements for cream-tinted body bg ("saturated brand color as body", "true off-white at chroma 0", "darker mid-tone tinted neutral"). Without them, the one-line warning left the model with no concrete alternative. Cross-provider validation showed the pattern matches historical behavior: Gemini benefits from prescriptive scaffold (+0.12 over off), Sonnet is roughly neutral (+0.01), GPT-5.5 slightly regresses (-0.11 matching the v3.1.0 pattern of -0.11). The skill has never been uniformly better than skill-off across providers; this is the closest achievable state without provider-specific rework. The structural improvements from the prior restructure stay (file merges, S-tier inlines, routing rule extension, reveal-animation safety rule). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>