refactor(skills): remove dated prompt patterns found by prompt audit

[ryzizub]

What

Remove dated prompt patterns from all 15 skills and the reviewer agent.

Why

The skills carry duplicated rules, failure-mode lists and incident-specific wording from eval tuning. Current models read those as extra constraints.

Changes

• Close the unclosed code fence in animations that broke its anti-patterns section
• Drop model: pins from every skill, so each answers on the session model
• Rewrite the five descriptions at the 1024-char cap as intent categories
• Replace repo paths in green-gate with skill-name references
• Remove duplicated rules, worked examples, and one editor-facing note
• Move the WCAG criterion tables to accessibility/references/wcag-criteria.md
• AGENTS.md, evals/README.md: pin count and structure updated

Evals

Commit 1 (two-arm, create-project + very-good-analysis-upgrade): every with-arm case ≥ 0.90, negative controls 1.00 in both arms. Other seven skills, with-arm: 41/49 at 1.00, rest match main.

Commit 2 (two-arm, the five rewritten descriptions): 36 cases, zero routing misses, 30 at 1.00, lowest 0.81. Other seven skills, with-arm: 37/45 at 1.00, lowest material-theming-refactors-hardcoded-widget at 0.75, which scores 0.79 with its old sonnet pin and fails the same two graders, so the pin is not the cause.

Pre-existing on main: dart-flutter-sdk-upgrade-reports-pub-get-conflict-instead-of-resolving scores 0.67. The model reads the pasted error as a toolchain mismatch. Needs its own fix.

🤖 Generated with Claude Code

[marcossevilla]

nice deletions!