Agents were interpreting "Add X to Y" as implicit permission to skip
brainstorming, TDD, or RED-GREEN-REFACTOR. Added general principle that
applies to ALL workflows:
"Your human partner's specific instructions describe WHAT to do, not HOW."
Then specific STOP for skill editing with required check-in.
Pressure tested and confirmed - agents now stop and check in before editing skills.
Succinct version (general principle ~5 lines + skill-specific ~12 lines) more
maintainable than repeating pattern in every workflow.
fc06ba4870