fix: improve skill optimization workflow precision

[shinpr]

Summary

• Wire user phrases and project-specific value into skill-creator description generation as mandatory inputs
• Make Progressive Disclosure Tier 1 (description quality) a grade gate — Tier 1 failure now blocks grade A/B
• Replace vague BP-001 "safety-critical/destructive" exception with strict 4-condition AND test + boundary examples
• Convert "Prohibited Actions" to positive-form "Operational Constraints" (BP-001 self-application)
• Add skill-creator modification mode for targeted section-level edits with changesSummary output
• Add WebSearch verification to skill-creator/skill-reviewer for knowledge cutoff protection
• Add patternExceptions output with per-condition evidence fields
• Add project-specific value validation (Round 2) and trigger phrase classification to create-skill
• Extend cross-skill detection to ~/.claude/skills/ in all commands and agents

Test plan

• Run /create-skill and verify 4-round collection flow (project-specific value in Round 2, phrase classification in Round 3, practical artifacts in Round 4)
• Verify skill-creator receives user phrases and project-specific value in its input
• Verify skill-reviewer grades C when description fails Tier 1 quality checklist
• Verify skill-reviewer outputs patternExceptions (not patternIssues) for legitimate BP-001 exceptions
• Run /refine-skill and verify it routes through skill-creator modification mode
• Verify EN/JA parity across all 14 changed files

🤖 Generated with Claude Code