Unfortunately, my experience with DeepSeek V3 API showed simple negative lists were disappointing.

• XML tags: Strongest adherence.
• Persona: Often results in cliches. Use XML to keep character voices consistent.

Stumbled upon this because I was also nervous about wasting credits. Using TypingMind Desktop App helped me test small prompts safely. Tbh, just giving the model one solid paragraph describing the tone's a decent option. It works better than those massive GitHub prompts that'll make the prose feel stiff. Sticking to simple instructions and setting temperature to 0.7 kept everything fast for me.

^ This. Also, been messing around with the API for a while now and honestly, you gotta be careful not to choke the model with too much junk. I agree with the XML tags, they really help keep the model focused, but dont overdo the length or the latency will start creeping up on you. I would suggest keeping your system prompt lean so you dont burn through those credits too fast. If youre doing this for a contest, maybe look at using Cursor AI Code Editor to manage your drafts and prompts in one place, it makes iterating way faster than just copy-pasting from a browser. Make sure to:

• Keep your style guidelines to a single, punchy paragraph.
• Use XML tags to define character voices, it just works better for consistency.
• Watch the temperature, keep it around 0.7 like mentioned before. Be careful about adding too many negative constraints, sometimes the model just ends up obsessing over what NOT to do instead of actually writing. Just give it a strong example of the prose you want instead.

Yep been there done that. Can confirm everything said above is spot on.