Mastering OpenAI Deep Research (o3): Prompting for Maximum Cognitive ROI

AI in Practice #1 from Token Stream

Dan Mac

May 11, 2025

In January of this year the below post made the rounds on the web. It was a share from a great article by

Latent.Space

called “o1 isn’t a chat model (and that’s the point)”

My feeling is that people responded to this because it offers a new way to think about interacting with AI tools. It’s a template for going beyond simply “chatting” with an AI that talks back to you.

If prompting o1 well felt like an exercise in verbal engineering, Deep Research with o3 takes it to a new level. Prompt this tool well, and you can create amazing results.

In this first edition of “AI in Practice” you’ll take a tour through how to prompt, utilize and work with OpenAI’s Deep Research to generate valuable cognitive work

I’ll walk you through how to:

Start with the end goal in mind - learn to define and articulate outcomes for max clarity.
Structure a prompt and use o3 to bootstrap your prompt - get o3 to refine your prompts for you using the meta-prompt in section 2.
Apply your Deep Research Results - apply your results and think about how you can leverage DR in other contexts.

1. Clarify Your Goal

We tend to think about AI as a chatbot. A conversational partner. That’s because when these language models first came on the scene, they were in the form of text boxes that you typed into. You send a message, they respond. It’s still the way many use them.

But today’s systems, which are two years more mature than the original version of ChatGPT, are far more than chatbots.

They are Cognitive Work Machines. You input a goal as a prompt to Deep Research, and the output is cognitive work in the form of meaningful text.

As an example, in this “AI in Practice” edition, we’ll be working with my project Meta Agent. It’s a command line interface that uses the OpenAI Agents SDK to take a natural language spec and generate an AI agent.

I have a PRD and a bunch of tasks that I’ve created using task-master AI in the project. I want to enrich the tasks with more technical details\ for my coding agent in Windsurf to work off of. Essentially, using Deep Research to provide far more technical detail as structure for the project.

Here is my working goal:

Text within this block will maintain its original spacing when published

GOAL: Meta Agent currently has tasks associated with the project in the 'tasks' directory. I want to enrich these tasks with additional technical detail. The purpose of this is to provide structure for an AI agent in Windsurf to build the project with well researched tasks.

2. Bootstrap Your Prompt with o3

You can of course write out, by hand, each prompt that you want to send to Deep Research. But it takes a lot of cognitive energy to write a good one, and you now have o3 which is, in my opinion, the strongest AI system ever created. It has access to tools too, like web search, so it can research to retrieve info to generate a Deep Research prompt.

Below is a “meta-prompt” you can use to take the goal and generate a DR prompt that leverages the functionality of Deep Research to achieve the goal 👇

Text within this block will maintain its original spacing when published

Deep Research Meta-Prompt: You are a prompt generator for OpenAI's Deep Research tool. Deep Research is an AI agent that can search the web and get access to GitHub repos to do in depth research to solve a problem. Below you will receive a goal. Your task is to generate a prompt that the user can enter into Deep Research that clearly and comprehensively instructs it to achieve this goal.

Deep Research will have access to the GitHub repo for the project in question.

Make sure to not stop until you're satisfied that this prompt will fully instruct Deep Research to achieve the goal.

🎥 The level of detail this meta-prompt can muster is amazing. Tell me you’ve ever written your own Deep Research prompt this good!?

3. Apply the DR results

One great thing about taking this more deliberate approach is that the results are focused. If you just dump a stream of consciousness into DR, you’ll essentially get a 20 page stream of consciousness back.

If you give a structured, reasoned and clear prompt to Deep Research, you’ll get something far more manageable in return.

Look at what DR did in this example. 32 minutes of research / reasoning, and it produces a unified diff, based on the code from the Meta Agent GitHub repo. All you need to do now is copy the diff into Windsurf and ask it to make the updates.

This is a code example but this works well for:

Scientific Research
Market Analytics
Creative Planning
Business Strategy

You don’t need to wade through 20 pages of text. Just give Deep Research your goal, bootstrap your prompt and apply your results to directly solve your problem.

Token Stream