How to craft the perfect AI prompt: the o1 prompt method by Greg Brockman

What is an o1 prompt

Unlike traditional queries that rely on associative connections between words, o1 queries are designed to activate the AI's logical capabilities. This approach allows the system not only to understand the essence of the question but also to analyze it more deeply.

Brockman emphasizes that the key to an effective o1 query is structure and specificity. A clear sequence—context, instructions, and expected outcome—enables the system to provide answers that older models cannot.

A good o1 query includes three key components:

1. Context — provide the system with the necessary background information. For example, if you are asking for travel advice, specify the location, dates, and preferences.

2. Instructions — clearly formulate the task and constraints. Need it to be brief? Mention it. Don't want deviations from the topic? Say so directly.

3. Output requirements — specify the format and style of the response: a list, a narrative, step-by-step instructions.

Step-by-step guide

A good result from AI is not a coincidence but the outcome of a well-constructed query. Here is how Brockman suggests acting:

1. Set a clear context
Give the system a starting point. The more specific you are, the higher the accuracy of the result.

Example: "I need routes for medium-length hikes within a two-hour drive from San Francisco. We have already done most of the local trails, so we are looking for something new, preferably with an ocean view".

2. Provide clear instructions
Formulate the task as specifically as possible.

Example: "Create a list of three routes. For each, include: the AllTrails name, start and end points, distance, driving time, hike duration, and what makes it unique".

3. Define the output format
Clarify exactly how the response should look: bulleted list, numbered list, tone of communication.

Example: "Make a list of three items with brief descriptions. Tone — friendly, informative".

4. Specify what to avoid
Warn against what you do not want to see in the response: unreliable data, well-known routes, speculation.

Example: "Ensure that all routes exist on AllTrails. Avoid well-known ones, such as Mount Tam. Verify the accuracy of distance data".

5. Add a "context block" if necessary
If the query relates to personal preferences, experience, or unusual conditions — mention this at the end.

Example: "For reference: my girlfriend and I often go hiking. We love routes with ocean views and food at the end. This will be our last outing before we part ways, so I want something special".

Why this works

A well-structured query gives the model a clear action plan. The fewer assumptions required, the better the result. This approach reduces the likelihood of errors, fabricated facts, and irrelevant answers.

Tips

— be specific: provide numbers, dates, constraints
— use action verbs: "explain", "list", "analyze"
— clarify the format: list, paragraph, headings
— define the style: formal, informal, concise
— don't be afraid to edit: queries can be improved.

Source: felloai.com.