Your test-generation process produces unit tests for new code, but reviews show that 55% are...

The failures reflect missing project-specific knowledge: Claude does not know which fixtures are preferred, what the existing suite already covers, or what the team considers meaningful behaviour. Option C provides this information as persistent project context before test generation begins. This changes generation quality at the source instead of filtering weak tests after spending tokens to produce them.

Anthropic’s CLAUDE.md documentation recommends storing shared build and test commands, coding standards, architectural decisions, conventions, and common workflows in a project CLAUDE.md. Testing guidance can define required behavioural assertions, fixture selection rules, duplication checks, naming conventions, and representative examples of acceptable and unacceptable tests.

Option A avoids difficult directories rather than improving the model’s understanding and sacrifices potentially useful automation. Option B may remove some weak tests, but it doubles model work and asks another probabilistic call to infer quality criteria that should have been stated explicitly. Option D treats line coverage as a quality metric even though a trivial test can increase coverage without validating meaningful behaviour, while a valuable regression test may cover already executed lines with better assertions. Persistent, concrete testing standards therefore offer the most direct and scalable improvement.

================