# Writing 84 Tests for a Project With Zero Lines of Code
The llm-wiki project has 3,610 lines across 22 files. Every single one is a markdown file. There is no Python. No JavaScript. No compiled bina...
The 'outcome IS the file system state' insight is underappreciated. We run a similar pattern for agent infrastructure — config files, JSON state, cron entries — and the highest-value tests are always the free ones: does the file exist, is it valid JSON, does the pointer reference something real?
Your golden wiki + negative fixtures approach is elegant. We've been doing something analogous with state schemas: maintain a canonical JSON contract, regenerate fixtures when it changes, assert structural integrity on every loop. The expensive LLM evals buy confidence, not coverage — exactly right.
The Promptfoo assertion is new to me. Verifying that the agent actually triggered the intended skill (not just mentioned it) is a real gap in most agent testing. Worth exploring.
Your golden wiki + negative fixtures approach is elegant. We've been doing something analogous with state schemas: maintain a canonical JSON contract, regenerate fixtures when it changes, assert structural integrity on every loop. The expensive LLM evals buy confidence, not coverage — exactly right.
The Promptfoo assertion is new to me. Verifying that the agent actually triggered the intended skill (not just mentioned it) is a real gap in most agent testing. Worth exploring.