OpenAI Cookbook — Official Prompting Guides

What it is

OpenAI Cookbook is the official collection of prompting guides and code examples maintained by OpenAI. It covers GPT-5.2, Codex, Meta Prompting, and the Realtime API, serving as the definitive reference for anyone working with OpenAI models.

This resource is aimed at developers, prompt engineers, and AI practitioners who want to move beyond trial-and-error prompting and adopt patterns validated by OpenAI themselves.

How it saves time or tokens

By following official best practices, you avoid common anti-patterns that waste tokens on retries and poorly structured prompts. The cookbook demonstrates techniques like system message structuring, few-shot formatting, and output constraining that reduce token consumption per request. Each guide distills lessons from OpenAI's internal testing, saving you hours of experimentation.

How to use

1. Browse the OpenAI Cookbook repository on GitHub to find the guide relevant to your model or use case.
2. Follow the step-by-step code examples, adapting them to your application's language and framework.
3. Apply the prompting patterns (few-shot, chain-of-thought, meta prompting) to your production prompts and measure the improvement in output quality.

Example

# Meta Prompting pattern from OpenAI Cookbook
system_prompt = '''
You are an expert prompt engineer.
Given a task description, generate an optimized prompt
that maximizes accuracy and minimizes token usage.
'''

response = client.chat.completions.create(
    model='gpt-4o',
    messages=[
        {'role': 'system', 'content': system_prompt},
        {'role': 'user', 'content': 'Write a prompt for summarizing legal documents'}
    ]
)
print(response.choices[0].message.content)

Related on TokRepo

• Prompt Library — Browse community-contributed prompts across models and use cases
• AI Tools for Coding — Developer tools that integrate LLM prompting into workflows

Common pitfalls

• Copying prompts verbatim without adapting them to your specific model version or API endpoint.
• Ignoring the distinction between system, user, and assistant roles when structuring multi-turn conversations.
• Over-engineering prompts with excessive instructions when a concise few-shot example would perform better.