mirror of
https://github.com/wondelai/skills.git
synced 2026-09-14 14:38:52 +08:00
mjaskolski
79bdb3a857
Both journeys were executed end to end in scratch projects with every constituent skill treated as uninstalled, to test the standalone path. Highest-severity defect, hit by both: references/artifact-templates.md shipped skeletons for only 2 of the 5-7 docs/ files each journey writes, while Operating Rule 7 orders creating missing files "from the full skeleton". The remaining headings live only in docs/ARTIFACT-REGISTRY.md, which is maintainer-facing and ships to nobody — so a standalone run invented them and silently dropped canonical sections. Both templates now carry every skeleton they write. conversion-optimization: - §1.4 leak-value formula was dimensionally wrong: raw lost users x LTV always ranks top-of-funnel first, defeating its own instruction. Now gap-to-benchmark x downstream rate x value, i.e. recoverable money. - §7 mandated a guardrail but sized only the primary metric; at funnel volumes the guardrail often cannot reach power. Added a sizing step with the monitored-not-gating fallback and a fast proxy. - Phase 1's GATE was unsatisfiable on a partly instrumented funnel; now gates on localizing the leak, with dark sub-steps awaiting-evidence. - Homes added for outputs that had none: Brand Script, Price Metric, and a Proposed message/CTA column so rewrites stop overwriting the before-state that Phase 7 measures against. - Phase 6 no longer demands checkout semantics in payment-free flows. architecture-optimization: - Rule 8 read literally auto-reverted every Phase 7-8 change, since timeouts, breakers and CI gates move no latency baseline; now scoped to changes aimed at a baseline. - Phase 1 fused two independent nets behind one gate; split so pinned behavior unlocks 2-4 and a baseline unlocks 5-6. - Phase 6 named an input Phase 5 never produces; now cites the Phase 1 profile and says to enable pg_stat_statements/auto_explain first. - Phases 5 and 6 may swap when the profile shows the database dominating — caching ahead of an N+1 fix is the skill's own named mistake. - Phase 1 now creates TECH-DEBT.md, which it was already told to write to. - methods.md gains pool sizing via Little's law, cache-stampede guidance, EXPLAIN (ANALYZE, BUFFERS), and composite index column order. Also: measured load supersedes the QPS estimate; benchmarks added for the funnel steps that had none; scope qualifiers so "every list endpoint" and "every outbound call" mean the in-scope flows; tracker records fallback vs skills-installed mode and what awaiting-evidence is waiting for; optional phases state their standalone method instead of implying one. create-website now routes to conversion-optimization (the only sibling border that was one-way). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>