CEO, AI Lounge & SaasRise
Recursive Self-Improvement Inside Your Software Factory
Three people at StrongDM ship software that no human writes or reviews. Dan Shapiro calls that level 5 of AI coding, the dark factory: specs go in, software comes out.
What makes a factory like that get better is a loop called recursive self-improvement (RSI): a system that improves itself, then uses the better version to improve again. The AI labs point it at their models. Claude wrote over 90% of Anthropic's code in September.
You can point the same loop at your software factory:
- Capture every run: the task, the tests, and every edit a human made.
- Score it, with code, a person, or another model as the judge.
- Read 20 runs and name the pattern in the low scores.
- Write one small fix back into a skill, a prompt or a test.
Two guardrails: a person approves every change to shared skills, and you measure the cost of each accepted change, failed attempts included.
Comment LOOPS and I'll send you how to set up a verification loop in Claude Code.
Videos






Comments1
The feedback loop makes sense, Ryan. The harder engineering problem is proving each improvement is actually an improvement. Better benchmark scores don't necessarily mean safer software. You need regression isolation, reproducible evidence, independent verification, and controlled promotion. Otherwise, recursive self-improvement can become recursive defect propagation.
1 comment
The feedback loop makes sense, Ryan. The harder engineering problem is proving each improvement is actually an improvement. Better benchmark scores don't necessarily mean safer software. You need regression isolation, reproducible evidence, independent verification, and controlled promotion. Otherwise, recursive self-improvement can become recursive defect propagation.