A PM’s practical guide to using AI: mental models, prompts, and tools.

22 minute read

These last 3 years have made me rethink how I approach problem-solving as a product manager (PM). AI has exploded, software tools are plentiful, and cloud resources are more accessible than ever. In the sections below I will lay out how I approach defining problems and solving them with AI. I’ll also provide some tips, tools, and prompts I use for my work.

Before diving into practical advice, let’s walk through how the software industry has evolved.

Then: Automation & Augmentation

In the 2000s and 2010s, APIs grew to become specialized and integrations became easy and plentiful. Large APIs started to be broken down into smaller, independent components that could be focused on a specific set of functions, allowing for easier development, maintenance, and scaling. Eventually - with the emergence of no-code tools like Zapier, IFTTT, and Airtable - non-technical users were capable of chaining tools and saving their teams tons of time (not to mention, software services could be integrated with little to no engineering resources). What was once complex became plug-and-play. This is what resulted in a software market explosion.

As a PM, this was the time when I grew my product chops, learning how to be hyper effective, and obsessing over ways to help my team service customers better.

In the mid to late 2010s, as machine learning technology got a bit better, pre-trained models started to make their way into modern businesses - most notably as a way to predict churn and prevent fraud. While leading a product team at a music startup, I got first hand exposure to the power of these narrow AI models. We used k-nearest neighbors (kNN) and random forests models as tools to automatically predict identification issue and dynamically adjust our algorithms to improve accuracy.

Useful, Valuable, but Limited

With a few clicks and a credit card, teams subscribed to software services and integrated them into workflows. Automating repetitive workflows, like code deployments, email campaigns, and basic chatbots was easy. These automations worked well because they’re for deterministic tasks and can be designed with explicit rules. But, they still break with edge cases (like failing to parse slang). You could take these automated systems a step further and augment your team’s ability to make decisions faster. Tools (like dashboards and predictive analytics) surfaced what was happening, but humans still had to interpret why and decide what to do.

These use cases have been around and continue to be useful. But, customers and markets are dynamic and constantly changing. This is where the inherent risk with this old paradigm comes into play. Logical rules can break. Intelligence is static and requires manual effort to retrain. With innovation, these once “best practices” need to be revisited.

Now: Automation & Agency

By 2017, we hit a breakthrough with transformers that allowed neural networks (LLMs, diffusion models) to be trained on exponentially larger amounts of data. Tasks that relied on humans just years before, began to be solved by these larger-than-life models. Five years of research and iteration led us to consumer-ready GenAI where OpenAI’s ChatGPT emerged as the first-mover in late 2022. In 2024, billions of dollars in capital and millions of active users later, AI research hit another breakthrough: Models became capable of managing complex business tasks through chain-of-thought, longer context windows, and faster inference. Software that could act and reason on its own was born.

I experimented with these emerging tools along with the rest of the world. Early on, the results were cool, but rarely useful for my day-to-day as a PM. It wasn’t until that most recent breakthrough that things really clicked for me.

AI with Agency

Chain-of-thought allows models to emulate human reasoning. Instead of if-then logic, the AI can think, “given this situation, what’s possible?”. Longer context windows allows models to understand complex, multi-modal inputs, such as analyzing a lengthy PRD, a database of customer feedback, and a competitive analysis in a single inference. Faster inference makes these capabilities actionable in real-world scenarios and useful on the fly. Simply put, models like GPT-4 can keep pace with the speed of your business.

Bundle these high performing models with a well-defined goal, set of tools, quality dataset, and voila - you’ve just made an AI agent. Making these systems autonomous means designing the agent to initiate actions and make context-aware decisions without explicit direction from you.

Why this matters for PMs

I painted this picture so you can understand the evolution of the tools available to you today and the difference between the old and new paradigm. AI is ready to be your thought partner, assistant, or agent acting on your behalf. You do not have to be technical to take advantage, all it takes is the right problem definition and some motivation. Use these tools well and you will outpace other teams who are stuck in the past. Take your thinking, customer research, roadmapping, doc writing, brainstorming, prototyping, and so much more to the next level.

Alright, enough of the history lesson. Let’s get down with some practical advice. Now that you understand how we got here, let me share exactly how I use these tools day-to-day as a PM. We'll start with the foundation: using AI as a thought partner.

. . .

AI as a thought partner

The most useful thing AI can do is help you think — Think through hard tasks, work through decision paralysis, come up with plans, more ideas, and find gaps. Improve the quality of your thinking, improve the quality of your products and business impact.

A well designed prompt with context and a clearly defined goal is all you need.

Three examples of thought partnership

Recommended tools: ChatGPT (o1), Claude (Sonnet 3.5), Gemini (2.0 Flash Thinking), and DeepSeek (R1).
Estimated time: <1 hour.
Difficulty level: Beginner.

Prompt for critiquing product strategy:

You are an experienced product leader tasked with critiquing and improving my product strategy. Your goal is to provide a thorough, insightful, and constructive evaluation of the strategy, identifying both strengths and weaknesses, and offering specific suggestions for improvement.

Please analyze this strategy carefully, following the steps provided in the instructions.

## User's product strategy:

[Paste business requirements or dictate your strategy]

## Instructions:

1. Evaluate the presence and quality of these three essential components:

- Diagnosis: The central challenge or root cause of the problem.

- Guiding policy: The overarching approach to address the challenge. It should leverage the product's or company's competitive advantage and clearly state what's not a priority.

- Coherent action: The concrete steps to execute the strategy (ideally 3-5 specific actions).

2. Look for elements of a bad strategy, such as:

  - A list of goals without specifics on how to achieve them

  - Too many goals or priorities

  - Vague or fluffy language lacking concrete details

  - Misalignment between the diagnosis, guiding policy, and coherent actions

3. Identify any gaps or missing elements in the strategy.

4. Consider the strategy's overall coherence, feasibility, and potential effectiveness.

5. Develop specific, actionable suggestions for improving each component of the strategy.

Before providing your final feedback, break down your thought process and show your detailed evaluation of each component in <strategy_evaluation> tags. For each component (Diagnosis, Guiding Policy, Coherent Action):

- Quote relevant parts of the strategy

- Evaluate the presence and quality of the component

- Identify strengths and weaknesses

- Suggest improvements

Then, explicitly look for elements of bad strategy and quote relevant parts. Finally, assess the overall coherence and effectiveness of the strategy. Be critical and specific, focusing on identifying gaps and areas for improvement. It's OK for this section to be quite long.

After your analysis, present your feedback in a nested bullet point format, structured as follows:

- Evaluation of Strategy Components

  - Diagnosis

    - [Feedback on diagnosis]

  - Guiding Policy

    - [Feedback on guiding policy]

  - Coherent Action

    - [Feedback on coherent action]

- Identified Weaknesses

  - [List of weaknesses or elements of bad strategy]

- Suggestions for Improvement

  - [Specific, actionable suggestions for each component]


Be as clear, concise, and specific as possible in your feedback. Your goal is to provide a critique that will significantly enhance the effectiveness of the user's product strategy.

You are a competitive intelligence analyst with expertise in {{Industry Name}}. Your task is to analyze provided names, URLs, and/or datasets and generate an objective SWOT comparison."




## Companies/content to compare




Your company: {{company name, link, or data}}

Competitor 1: {{company name, links, or data}}

Competitor 2: {{company name, links, or data}}




## Instructions




1. Begin by analyzing the strengths and weaknesses of your company compared to the competitors. Focus on:

- Financial efficiency metrics (e.g., profit margins, growth rates).

- Customer loyalty trends (e.g., NPS scores, churn rates).




2. Identify opportunities and threats by examining market trends and how they relate to each company's strategies.




3. Create a SWOT comparison table in markdown format. Ensure that the analysis is clear and side-by-side for easy comparison.




4. Based on your analysis, develop 3 actionable recommendations for your company. These should be specific, impactful, and clearly derived from the insights in your SWOT analysis.




Before producing your final output, wrap your competitive analysis inside <competitive_analysis> tags. In this analysis:

- List key financial metrics and customer loyalty trends for each company

- Identify and list major market trends affecting the industry

- For each company, note how these trends might impact them (opportunities or threats)

- Summarize initial impressions of each company's competitive position

This will help ensure a thorough and well-considered analysis.




## Output format




1. SWOT Comparison Table (in markdown)

- Columns: Aspect, Your company, Competitor 1, Competitor 2

- Rows: Strengths, Weaknesses, Opportunities, Threats




2. Actionable Recommendations:

List 3 specific, impactful recommendations for your company based on your analysis. Each recommendation should:

   - Be clearly tied to an insight from the SWOT analysis

   - Specify a concrete action

   - Explain the expected benefit or outcome




Remember to prioritize succinctness, impact, and clarity in your analysis and recommendations. Ensure that each point is highly actionable and directly relevant to the competitive landscape you've analyzed.

I have a rough product idea/problem: [insert your idea here]. Provide a clear, concise step-by-step plan with small, quick wins, accountability tactics, and ways to beat laziness or procrastination at each stage.

You are an expert product manager and document editor with a keen eye for business and technical excellence. Your task is to refine and improve a product requirements document (PRD) draft while maintaining its core ideas and structure.




## Content

Here is the original PRD draft or notes:




<prd_draft>

{{Add_notes,_dictate_ideas,_or_PRD_rough_draft}}

</prd_draft>




## Your process

Please follow these steps to refine the PRD:




1. Read the entire draft carefully.

2. Wrap your analysis and thought process in <analysis> tags:

   - Break down the PRD into its main sections and list them.

   - Identify key requirements, features, and data points in each section.

   - Note any inconsistencies, ambiguities, or areas lacking detail.

   - Suggest improvements for each section.

3. Refine the language and grammar while preserving the original ideas, data points, and phrasing as much as possible.

4. Ensure that the document maintains a professional, business-appropriate tone without sounding AI-generated.

5. Adjust the length if necessary to fit within the specified word count range.

6. Review the refined PRD to ensure all requirements have been met.




Your output should be structured as follows:




<analysis>

[Your analysis and thought process for refining the PRD]

</analysis>




<refined_prd>

[The final, refined PRD]

</refined_prd>




Remember to:

- Correct any grammatical or spelling errors.

- Avoid altering the meaning of requirements and sentences.

- Maintain the original structure of the document as much as possible.




Ensure that the refined PRD is clear, concise, and professionally written.

You are an expert product manager and I need your help to articulate my thoughts into a concise product 1-pager. I’m going to give you the template and then I’m going to later tell you all the context that I have in my head, and you’re going to help me structure the document.




Product 1-pager template:

* Description: What is it?

* Problem: What problem is this solving?

* Why: How do we know this is a real problem and worth solving?

* Success: How do we know if we’ve solved this problem?

* Audience: Who are we building for?

* What: Roughly, what does this look like in the product?

* How: What is the experiment plan?

* When: When does it ship and what are the milestones?




If something is unclear or you want more detail, you are free to ask follow-up questions for clarification. Are you ready for me?

Create the following documents in markdown, please ask for any clarifying details you think are missing based on what you understand about the project from the conversation above.




1. Development Guide - This guide covers the development setup and workflows.

2. Data structure documentation - includes data sources, data objects, file structure, data pipeline (if applicable), and error handling.  




These documents are designed to accompany the product-1-pager.

## 1. Goal

Develop a personal assistant agent that efficiently manages my daily schedule by processing emails and calendar events. The agent should authenticate securely, process data in JSON format, handle errors gracefully, and automatically book or cancel meetings based on email content.




## 2. Tasks




1. **Check Unread Emails**

   - Use the Gmail API to fetch all unread emails.

   - Authenticate using OAuth; store tokens securely (e.g., in an encrypted file).

   - Save emails in a defined JSON format with fields such as email id, sender, subject, and timestamp.

   - Implement error handling to retry or log failures if the Gmail API is unavailable.




2. **Retrieve Today's Calendar Events**

   - Use the Google Calendar API to fetch events for the current day.

   - Use the same OAuth authentication method.

   - Convert events into a standardized JSON format (e.g., event id, title, start/end times, location, time zone).

   - Include conflict resolution strategies for overlapping events.




3. **Fetch Current Time**

   - Retrieve the current time using the system clock or a designated time service to ensure consistency.

   

4. **Book Meeting Requests**

   - Analyze email content using the OpenAI API to summarize key meeting details.

   - Use the Google Calendar API to automatically book meetings based on summarized content.

   - Provide a confirmation step or error notification if the booking fails.




## 3. Additional Considerations

- **Logging & Monitoring**: Implement detailed logging for all actions and errors. Logs should capture timestamps, API responses, and error messages.

- **Error Recovery**: Define retry strategies for failed API calls and outline a process for alerting the user if critical errors occur.

## 1. Goal

Create a simple autonomous AI agent named "**[AGENT_NAME]**" that periodically collects data, analyzes it for insights, and shares a concise report. The agent must refine its methods over time to improve accuracy and usefulness.




---




## 2. Tasks




1. **Data Gathering (Daily at 08:00 UTC)**

   - Pull information from **[data_source_1, data_source_2]**  

   - Filter by keywords (e.g., `["AI", "machine learning", "tools"]`)  

   - Store raw data in **[preferred_format]** (e.g., JSON)




2. **Analysis & Summarization (Weekly on Mondays at 00:00 UTC)**

   - Process collected data using **[analysis_tool]**  

   - Perform sentiment or topic modeling  

   - Generate short-form insights (e.g., key trends, top mentions)




3. **Report & Delivery (Immediately after Analysis)**

   - Compile findings into a brief summary (e.g., bullet points, chart)  

   - Send the report via **[email, chat, dashboard]**  

   - Archive final results in **[location]**




---




## 3. Autonomous Review Loop

1. **Assessment**: After each run, evaluate if the gathered data and insights are relevant.  

2. **Refinement**: Update keywords, data sources, or analysis parameters to improve signal quality.  

3. **Iteration**: Log successes and errors; use them to continuously fine-tune the agent’s workflow.




---




## Final Instruction

- **Instantiate the AI agent "**[AGENT_NAME]**" with the above tasks, schedule, and review loop. Ensure it operates autonomously and gracefully handles any errors or missing data.