Effective Prompts for Extended Thinking
Keep prompts simple, avoid step-by-step instructions, trust the model to reason.
Prompting Reasoning Models Is Different
The prompting intuitions you built for standard models — chain-of-thought, step-by-step instructions, few-shot examples — often work against you with reasoning models.
Reasoning models are already doing sophisticated internal reasoning. Telling them exactly how to think can interfere with that process. The best prompts for reasoning models are simpler and more direct than prompts for standard models.
Don't Instruct Step-by-Step
With standard models, you write: "Think step by step. First consider X, then consider Y, finally conclude Z." This scaffolding helps because standard models don't do this automatically.
With reasoning models, this scaffolding can constrain the model's internal reasoning into a suboptimal path. Instead, state the problem clearly and let the model determine how to reason about it.
# Standard model: needs scaffolding
STANDARD_PROMPT = (
'Let us think step by step.\n'
'First, identify the variables.\n'
'Then, set up the equation.\n'
'Then, solve for x.\n'
'Finally, verify your answer.\n\n'
'Problem: If 3x + 7 = 22, what is x?'
)
# Reasoning model: just state the problem clearly
REASONING_PROMPT = (
'Solve: If 3x + 7 = 22, what is x?'
# The model handles the step-by-step internally
)
# Both produce correct answers; the reasoning model prompt is simpler
print('Reasoning model prefers the cleaner prompt')All lessons in this course
- How Reasoning Models Differ
- Effective Prompts for Extended Thinking
- When to Use Reasoning vs Standard Models
- Cost and Latency Tradeoffs