Reasoning models perform their own internal chain-of-thought, so the prompting strategy differs from standard models. Keep prompts concise and direct. Avoid telling the model how to think; instead, specify what you want clearly.
Reasoning Model: Direct Problem Solving
For reasoning models, less is more. State the problem clearly and let the model's built-in reasoning process work. Code Fragment I.8.1 below puts this into practice.
// Reasoning Direct user template
// Replace {{placeholders}} with your actual values before sending
{{problem_statement}}
Provide your final answer at the end, clearly labeled.// Reasoning Analysis user template
// Replace {{placeholders}} with your actual values before sending
Analyze the following {{subject}}:
{{content}}
Provide your analysis covering:
1. {{aspect_1}}
2. {{aspect_2}}
3. {{aspect_3}}
Conclude with a clear recommendation and confidence level (high/medium/low).Do NOT use chain-of-thought instructions ("think step by step") with reasoning models. They already think step by step internally. Adding such instructions can actually degrade performance by interfering with the model's native reasoning process. Instead, focus on making the problem specification clear and complete.
Reasoning Model: Complex Analysis
For multi-faceted analysis tasks, structure the desired output format rather than the reasoning process.
Reasoning models use "thinking tokens" that count toward token usage but are not always visible in the API response. This can make them 5 to 20 times more expensive per query than standard models. Use reasoning models selectively for tasks that genuinely require deep reasoning (math, complex logic, multi-step planning) and use standard models for simpler tasks like classification, extraction, or summarization.
Quick Reference: Choosing a Template
| Task | Template | Recommended Temperature | Model Type |
|---|---|---|---|
| Classification | Zero-shot or Few-shot | 0.0 | Standard |
| Summarization | Structured Summary | 0.0 to 0.3 | Standard |
| Data Extraction | JSON Extraction | 0.0 | Standard (with JSON mode) |
| Q&A (RAG) | RAG Q&A | 0.0 to 0.2 | Standard |
| Code Generation | Code with Spec | 0.0 to 0.2 | Standard or Reasoning |
| Math / Logic | Chain-of-Thought | 0.0 (standard) or default (reasoning) | Reasoning preferred |
| Creative Writing | Custom system prompt | 0.7 to 1.0 | Standard |
| Complex Analysis | Reasoning Analysis | Default (1.0) | Reasoning |
The best prompt is the one you have tested on your actual data. These templates are starting points, not finished products. Build an evaluation set of 20 to 50 representative inputs, measure output quality systematically, and iterate. Small wording changes can produce large quality differences, so treat prompt development as an empirical process, not a one-shot exercise.