- Step 2: Add "Match specificity to task risk" framework (high/medium/low freedom) - Step 6: Add refinement filter to avoid duplicate/speculative improvements These 5 lines help Claude choose between agentic (text instructions) vs scripted (exact code) approaches based on task characteristics. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>