prompt-engineer

Prompt optimization across LLMs involves applying core techniques—XML tags, role prompting, clear instructions, multishot examples, and chain-of-thought reasoning—tailored to each model's strengths. Use model-specific guides to select the best approach for Claude Opus, Sonnet, Haiku, GPT, or Gemini, then refine based on output quality and task requirements.

Prompt migration requires understanding each model's capabilities and constraints. Start with your original prompt, apply model-specific optimization techniques from the reference materials, test outputs across target models, and adjust for performance differences. Use the comparison guides to identify which model best fits your task before finalizing.

Chain-of-thought reasoning breaks down complex reasoning within a single prompt to improve accuracy on logical tasks. Prompt chaining sequences multiple prompts where output from one feeds into the next, useful for multi-step workflows. Choose chaining for decomposed workflows and chain-of-thought for enhanced reasoning within one interaction.

XML tags are a core prompt engineering technique supported across Claude, GPT, and Gemini models. They structure input and guide model behavior effectively. Check the model-specific reference guides to confirm syntax and best practices for your target model, as formatting conventions may vary slightly.

Model selection depends on task requirements, latency, cost, and capability fit. Use the model comparison reference material to evaluate Claude Opus, Sonnet, Haiku, GPT 5.1, and Gemini Pro 3.0 across dimensions relevant to your use case. Test with multishot examples and optimization techniques to validate performance before committing to one model.

Inconsistent performance often stems from model-specific differences in instruction interpretation, context handling, or capability gaps. Troubleshoot by reviewing the optimization checklist, applying role prompting and structured formatting techniques, and testing with shorter or clearer instructions. Compare outputs across models to identify which technique adjustments improve results.