How to Prompt Reasoning Models Effectively
From the course OpenAI o-Series: Reasoning Models Masterclass
Prompting reasoning models is fundamentally different from prompting standard models. The key insight: less instruction often produces better results.
The Counter-Intuitive Rule
With GPT-4o, you might write:
"Solve this step by step. First identify the variables. Then set up equations. Then solve each equation. Then verify your answer."
With o-series models, just write:
"Solve this: [problem]"
The model's internal reasoning is better than most human-prescribed step sequences.
Prompting Principles
- State the problem clearly — Focus on what you want solved, not how to solve it
- Provide all necessary context — Include all data, constraints, and requirements
- Specify output format — Tell it what the answer should look like
- Don't micro-manage reasoning — Let the model choose its approach
- Use system prompts sparingly — Short, focused system prompts work best
Good vs Bad Prompts
❌ Bad (over-instructed):
"I need you to solve this math problem. Please follow these steps: 1) Read the problem carefully, 2) Identify all given values, 3) Determine what formula to use, 4) Substitute values, 5) Calculate, 6) Verify your answer by plugging it back in."
✅ Good (clear problem, minimal instruction):
"What is the maximum area of a rectangle inscribed in a circle of radius 5? Give the exact answer and a decimal approximation."
Advanced Prompting Techniques
Constraint specification:
"Design a database schema for a library system. Constraints: must handle multi-branch locations, support digital and physical media, track borrowing history, and handle waitlists. Optimize for read-heavy workloads."
Multi-part analysis:
"Analyze this business proposal: [details]. Evaluate: (1) market viability, (2) financial projections, (3) competitive risks, (4) team capability, (5) overall recommendation with confidence level."
Verification requests:
"Solve this differential equation: [equation]. Then verify your solution by substituting it back into the original equation."
Temperature Settings
- Temperature 1 (default) — Best for most reasoning tasks
- Lower temperature (0-0.5) — For deterministic, reproducible results
- Higher temperature — Generally not recommended for reasoning
Handling Long Problems
For complex problems with lots of context:
- Put all data at the beginning of the prompt
- State the question clearly at the end
- Specify output constraints (format, length, detail level)
- Don't break the problem into artificial sub-steps
Common Mistakes
- Telling the model to "think step by step" — it already does this internally
- Providing worked examples — can actually constrain the model's reasoning
- Using chain-of-thought prompting techniques — designed for standard models
- Setting temperature to 0 for every task — limits exploration on open-ended problems
Key Takeaways
- •Less instruction often produces better results — let reasoning models choose their approach
- •Focus prompts on what to solve, not how to solve it
- •Don't use chain-of-thought prompting — the model reasons internally and better
- •Provide all context upfront, state the question clearly, specify output format
- •Use default temperature (1) for most reasoning tasks
Progress is kept in this browser only.