Prompt engineering

Root prompts/ contains task-specific system/user Markdown, a Pydantic-generated response schema, representative positive and counterexamples, and version guidance. Wheels bundle these files as package data. System messages never contain imported source. Context is JSON serialized.

assist hashes system+template and records the hash/version with provider/model/token metadata. When updating a prompt, update its version, examples and test expectations. Run pytest tests/test_llm.py for schema equality, template completeness, explicit evidence and hallucination rejection. Exact quotes must match original unredacted evidence after privacy redaction; a mismatch is rejected rather than approximated.