The methodology doc. Argues that prose specifications and test suites both fail to bind an AI agent over a long session, and lays out the mechanical rule set that does.
MAXWELL COLLINS
I'm writing a methodology called Domain-Rules-Driven Development. The argument: you can specify software in a form an AI agent cannot silently overrule, as long as the form is mechanical enough to check against source code at commit. I started writing it down after a 20-minute QA audit on March 30, 2026 turned up over 100 critical bugs in production from a codebase whose CI was green on 4,800 tests. The rule set replaced the test suite, with 42 rules locked so far. The count goes up when an incident forces a new one.
FEATURED
RECENT FIELD NOTES
- Why I Stopped Writing Tests First
Tests assert call-site values. User journeys are sequences. The gap is why test-first produces a specification with holes, and the six-stage pipeline that closes it.
- The 4,800 Fake Tests Weekend
The 20-minute QA audit that killed a passing test suite.
REFERENCE
The rule set has 8 entries on this site. The live rebuild has 42. Publishing lags the lock date by whatever it takes to remove the parts that identify the codebase.
Browse rules →