1. Analyze Original:
   • Note verbose instructions assuming model knowledge.
2. Rewrite:
   • Shorter: "Summarize this article in 3 bullets." vs. model-specific chains.
3. Test Across Models:

   prompts = ["Original", "Rewritten"]
   for p in prompts:
       response = model.generate(p)
       score = evaluate(response)
   

Custom scorer

4. **Iterate**: Use few-shot examples if needed.

## Example

GPT prompt: "Think step-by-step..." → Claude: "Analyze logically: [task]".

## Evaluation
- Metrics: ROUGE for tasks, human preference.
- Trade-offs: Brevity vs. guidance.

## Conclusion

Prompt rewriting ensures portability; always validate on target models.