Prompting across models
What changes when you move a prompt to another model.
DesignerDeveloper
What to re-check on every model
| Aspect | Action |
|---|---|
| Roles and slots | Put durable rules in the system or developer slot |
| Structure | Test tags, headings or plain labels |
| Verbosity | State length, tone and format limits |
| Reasoning models | Give goals and constraints, not step scripts |
| Small models | More explicit steps, a format, one or two examples |
| Examples | Start with few and varied; add only if tests improve |
| Output format | Prefer structured output or tools, then validate in code |
| Sampling | Read the docs; don't assume settings exist |
| Open-weight | Use the tokenizer's chat template |
| Languages | Evaluate each one you serve |
Process
- 1. Keep a portable prompt spec
- 2. Render per model
- 3. Run the same evaluation set
- 4. Fix the smallest thing per model
- 5. Record versions and reasons
- 6. Re-run on every model update
Pitfalls
- Copying prompts written for another model
- Conflicting instructions
- Examples that leak into answers
- Relying on undocumented quirks
- Testing in one language, shipping in five