Relevance Decay

Instructions are only effective for tasks they're semantically related to. A testing instruction has no effect on documentation tasks — the instruction must be relevant to the work at hand.

Antipatterns

• Mixing unrelated domains in one instruction. Writing "Use pytest fixtures and keep documentation concise" tries to address two unrelated task types. Neither domain gets strong coverage because the instruction dilutes itself across contexts.
• Generic instructions intended for all tasks. Writing "be thorough and careful" has near-zero effect on any specific task. Domain-specific vocabulary is required for the instruction to activate during relevant work.
• Assuming cross-domain transfer. A testing-specific instruction like "always write edge-case tests" will not make the model more thorough when writing documentation. Each domain needs its own instructions.

Pass / Fail

Pass

# Testing

Run `uv run pytest tests/ -v` before committing.
Use `@pytest.mark.parametrize` for multiple input cases.

# Documentation

Write `docs/*.md` for people, not agents.

Fail

# Quality

Be thorough and careful in all tasks.
Always produce high-quality output.
Double-check your work.

Fix

If you need the same behavior across diverse task types, write separate versions of the instruction with domain-specific vocabulary for each. "Use pytest fixtures for test setup" only works for testing tasks — if you also want consistent patterns in scripts, write a separate instruction naming script-relevant constructs.

Limitations

Measures semantic distance between instructions and task context. Cannot determine whether an instruction was intended to apply broadly or narrowly.