Three Loops: Retro, Macro, Meta

Most teams have one feedback loop. They do retrospectives. What happened this sprint? What went well? What didn't? This is the retro loop and it operates at the level of individual events. It's necessary and insufficient.

The retro loop operates at the conversation level. This lead went cold — why? The AI agent misclassified a response. The follow-up cadence was too aggressive. The building match was wrong. Each failure is a lesson. Each lesson becomes a rule. The CLAUDE.md grows by one line.

# Retro loop example — actual production rule

## Bottom-Up Analysis Law (ABSOLUTE — enforced by hook)
- NEVER run aggregate queries as your FIRST step
- ALWAYS sample 20+ individual records first
- Classifiers lie. Summary fields are stale.
- client_stage_progression.stage_name is garbage.

# This rule exists because I spent months making
# business decisions based on classifier fields
# that turned out to be LLM-generated freeform labels.

That's a retro loop output. I found a specific failure (garbage classifier field), extracted the lesson (don't trust summary fields), and encoded it as a rule (the bottom-up analysis law). Now a hook enforces it. The lesson persists even when I forget it.

The macro loop operates at the pattern level. Not "what happened in this conversation" but "what's happening across 30,000 conversations this month?" Are qualification rates trending down? Is a particular building generating complaints? Are follow-up cadences converting differently in Dallas than Chicago?

The macro loop found these patterns:

Chicago:
  44% of leads → wrong geographic match
  $280 gap between 2br listed vs actual
  68% of successful placements → different
  neighborhood than inquiry

Dallas:
  37% geographic waste (leads shown buildings
  outside comfort zone)
  Tour-to-close ratio 2x higher when building
  is within 3 miles of current ZIP

The macro loop requires data infrastructure that most organizations don't have because they're still manually reviewing individual cases. When you have 30,000 conversations a month, you can see patterns that are invisible at the individual level. The geographic waste pattern — 37% of Dallas leads being shown buildings outside their comfort zone — only emerged from aggregate analysis across thousands of conversations. No single conversation would reveal it.

The meta loop is the interesting one. It operates at the system level. Not "what's happening in conversations" and not "what patterns emerge across conversations" — but "is the system that generates conversations getting better?"

Meta loop questions:

- Is the CLAUDE.md getting more precise or more bloated?
- Are the hooks catching real problems or generating noise?
- Is session memory actually helping, or am I ignoring it?
- Are the red team agents finding new failure modes,
  or just re-finding old ones?
- Is the 1:1000 operating leverage ratio improving?

The meta loop is where you audit the auditors. The hooks prevent bad deploys — but are the hooks themselves correct? The CLAUDE.md encodes rules — but are the rules still relevant, or has the system evolved past them? The session memory persists context — but is it useful context, or noise that clutters the next session?

Most organizations never get past the retro loop. They review incidents, write postmortems, and move on. The best organizations build the macro loop — dashboards, analytics, trend detection. Almost nobody builds the meta loop, because it requires asking uncomfortable questions about whether your own processes are working.

I run all three loops simultaneously, on every business unit, continuously. The retro loop is the hooks and the CLAUDE.md rules. The macro loop is the data pipeline that aggregates across all conversations. The meta loop is the skill called /audit-claude-md that reviews the entire governance system for bloat, outdated rules, and content that should be moved or deleted.

The meta loop is also this essay. Writing about the system forces me to examine it from outside. Is the three-loop framework actually making things better? Or is it just adding complexity that makes me feel productive?

The honest answer: I don't always know. But the loops keep running whether I'm confident or not.