A principle emerging from agentic coding: verification steps that pass should produce no output; failures should dump everything. Simple idea, but it matters more than you'd expect when an LLM is the one reading the output.