Feedback Flywheel

Martin Fowler article on building compounding learning loops for teams using AI coding tools. The missing piece isn’t the tools — it’s the mechanism for effort to accumulate.

Key Claims

• Most teams plateau with AI tools — same prompting habits, same frustrations month after month
• Individual developer intuition about AI usage doesn’t transfer to the team
• Four types of signal from AI interactions:
  1. Context signal → feeds back into priming documents
  2. Instruction signal → feeds back into shared commands
  3. Workflow signal → feeds back into team playbooks
  4. Failure signal → feeds back into guardrails/anti-patterns
• Four cadences: after each session, at standup, at retrospective, periodically (quarterly)
• Measure: first-pass acceptance rate, iteration cycles, post-merge rework, principle alignment — NOT speed/lines generated

Takeaways

• The flywheel metaphor: each rotation leaves infrastructure better prepared for the next
• “Speed measures volume, not value. A fast output requiring extensive rework is rework with extra steps”
• This connects to the broader pattern of Compounding Knowledge — team learning about AI usage compounds like any knowledge
• Practical, lightweight — heaviest cadence is 5 minutes in an existing meeting

Connections

• Compounding Knowledge — the core principle at work
• Cognitive Debt — what happens without the flywheel
• How Copilot Makes Programmers Worse — the problem this addresses
• Coding Agents — the tools being discussed