prompt
FreePrompt engineering best practices for reasoning AI models
FreeFree tier
About prompt
A comprehensive prompt engineering guide for reasoning models (e.g., o1, o3, o4-mini, Claude 3.7/Sonnet 4.6, Gemini 2.0 Flash Thinking). It explains how these models differ from standard models due to their internal chain-of-thought reasoning, provides golden rules for effective prompting (such as stating goals clearly, using XML tags, and setting output constraints), and lists common anti-patterns to avoid (like explicit step-by-step instructions or excessive few-shot examples). The guide is part of the ai-boost/awesome-prompts open-source repository on GitHub.
Key Features
Explains the internal chain-of-thought mechanism of reasoning models
Provides golden rules for effective prompting
Lists specific do's (e.g., clear goal, context, XML tags) and don'ts (e.g., avoid 'think step by step', avoid prescribing reasoning path)
Covers models: o1, o3, o4-mini, Claude 3.7/Sonnet 4.6, Gemini 2.0 Flash Thinking
Includes references to official documentation from Anthropic, OpenAI, and Helicone
Open source and freely accessible on GitHub
Pros & Cons
Pros
- Comprehensive and actionable best practices
- Based on official sources and community experience
- Free and open source with no restrictions
- Clear structure with separate do/don't sections
Cons
- Only available as a text file, no interactive interface
- Limited to text-based guidance, no built-in prompt testing or validation
- May become outdated as reasoning models evolve
Best For
Improving prompts for complex reasoning tasksLearning best practices for prompting advanced AI modelsAvoiding common pitfalls when using reasoning modelsTraining or onboarding new prompt engineers
FAQ
What is the main difference between reasoning models and standard models?
Reasoning models run an internal chain-of-thought before producing their response, which is hidden from the user. This allows them to self-plan and excel at multi-step problems, but they are slower and more expensive per token.
Should I use 'think step by step' prompts with reasoning models?
No, explicit step-by-step instructions interfere with the model's native reasoning and degrade performance. Instead, state the goal clearly without prescribing the reasoning path.
Are few-shot examples recommended for reasoning models?
Few-shot examples often hurt performance because they constrain the model's natural reasoning path. Zero-shot prompting is typically more effective.
What models does this guide cover?
The guide covers o1, o3, o4-mini, Claude 3.7/Sonnet 4.6 adaptive thinking, and Gemini 2.0 Flash Thinking.