Evolution Protocol

How AI systems improve over time — session logging, quality feedback, the 3-strike rule, and quarterly reviews.


Session Logging

Every AI work session is logged with:

  • Task — what was requested
  • Routing — which agent or system handled it
  • Output quality — assessment of the result
  • What worked — patterns to reinforce
  • What did not work — patterns to correct
  • What was learned — new intelligence for the system

This log is the raw material for system improvement. Without it, the AI stays static. With it, the AI compounds.


Quality Feedback Loop

Positive Examples

Saved with WHY they worked. Not just "this was good" — what specifically made it good. This builds pattern recognition so the system can reproduce quality consistently.

Negative Examples

Saved with WHY they failed. Not just "this was wrong" — what specifically went wrong and what the correct output would have been. This prevents repetition of the same errors.

The 3-Strike Rule

If the same error occurs across three sessions, the problem is not the output — it is the foundation. A foundational file update is needed, not just a correction. Three strikes means the CLAUDE.md, positioning map, or operations map has a gap that needs filling.


Quarterly Review Checklist

Every quarter, review the full system against these questions:

  • Has the founder's voice evolved? Update CLAUDE.md.
  • Have offers changed? Update offer architecture and positioning.
  • Are there new competitors? Update market positioning map.
  • Have tools changed? Update business operations map.
  • Are any agents underperforming? Retrain or retire.
  • Is there knowledge ready to graduate from session logs to permanent docs?

The Compounding Effect

Timeline System Capability
Month 1 System sounds right. Can draft and process.
Month 3 Handles onboarding, proposals, content with minimal oversight.
Month 6 Anticipates needs based on accumulated intelligence.
Month 12 Significant operational load handled. Founders focus on strategy and high-touch work.

When to Expand

  • Add agents when a recurring workflow lacks coverage
  • Add SOPs after 3+ consistent executions of the same process
  • Prune when a component has gone unused for 2+ quarters

The system grows deliberately, not by default. Every addition must earn its place.



Sources

  • evolution-protocol.md — full protocol definition, logging templates, review checklists, expansion criteria