Prompts

Prompt assets in this directory are consumed by the runtime prompt loader in src/personal_llm/prompts/loader.py.

How these files are used

• system_professional.md: default guarded system prompt
• system_professional_strict.md: stronger professional-only behavior
• system_professional_relaxed.md: softer mixed-topic handling
• system_original_model.md: minimal repository steering
• refusal_redirect.md: default refusal and redirect wording
• refusal_redirect_relaxed.md: softer refusal wording
• refusal_disabled.md: near-original-model mode
• legal_finance_disclaimer.md: reusable caution text for regulated domains
• topic_router.md: topic-classification support prompt

Relationship to the persona file

The user-facing voice should start in knowledge/persona.md, then be translated here into runtime prompts.

Best practice:

1. define the voice in knowledge/persona.md
2. keep these prompt files aligned with that voice
3. create SFT examples that reinforce the same behavior
4. compare prompt-only behavior against LoRA behavior during evaluation