The Method
Not “how to use AI” — how to systematize your judgment so AI makes you more effective, not more reckless.
Current perspective · September 6, 2026
The guard needs evidence too
Later cases exposed checks whose output was not consumed, controls left behind by a service migration, and green tests that needed a deliberate failure probe. My standard is to inspect the actual reader and rejection path. A stored rule, an executed hook, and an enforced decision are different facts.
Amdahl's Law x AI
Why your judgment is your most valuable asset
AI accelerates the parallelizable part — writing code, searching, generating content. Platforms will commoditize this. The real value is in the serial bottleneck: your domain judgment, quality standards, strategic decisions. The stronger AI gets, the more your judgment matters.
Question-Driven Governance
Don't trust unquestioned output
732 tests passed. 0 failures. I was ready to celebrate — and my AI partner said 'Wait. Last time we had all-green tests too, and production broke.' It remembered an incident I'd forgotten. This isn't a tool. It's a partner with memory and the discipline to push back.
Four-Layer Memory Architecture
Context with an explicit handoff
Asset Layer (Obsidian) stores knowledge and decisions. Runtime Layer (CLAUDE.md) holds active rules. Enforcement Layer (Hooks) runs mandatory checks in code. Training Layer (Memory) captures feedback that graduates into rules. Cross-session, cross-tool, cross-project.
Self-Evolving Enforcement Layer
Rules that live in code, not documents
Hooks and scoped scanners check known failure paths. A platform may ignore an output, a workload may move outside a guard, and a text scan may miss an image. I check the consumer and exercise rejection paths, with retirement conditions for rules that stop earning their place.
Battle-Tested Across Real Products
Not theory — documented across real production
Production incidents, controlled checks, and issues caught before delivery document how the workflow changed. Each case separates the observed result from what remains untested.