How I AI: 3 Game-Changing Workflows for Product Managers

Even after a two-month paternity leave, Amir Klein, an AI product manager at monday.com, ranked fourth out of 90 product managers for his AI tool usage. An internal company leaderboard confirmed he was a power user, and he credits a specific system for his efficiency.

He has become so accustomed to the setup that he jokes he is now incapable of doing his job without AI, primarily because it solves the constant context switching that defines a product manager’s day.

Instead of relying on his own memory, Amir creates a separate Claude or ChatGPT workspace for each initiative. Each one acts as a dedicated "second brain" that holds the approved documents, decisions, and research for that project, allowing him to return to the work without reconstructing all of the context from scratch.

In this episode of How I AI, Amir demonstrates three of his most effective workflows: a project workspace informed by public customer conversations, a custom GPT for concise writing, and mock product interviews in GPT Voice Mode.

He shares the setup, prompts, and limitations of each workflow, including the manual checks he performs before trusting an analysis or sending an AI-edited message.

Scraping public forums to build a project-specific 'second brain'

Amir’s project workspaces reduce the context he has to hold in his own memory. Each contains relevant documents, discussions, research, and decisions for one initiative, so he can answer a stakeholder question on the spot without having to start his research all over again.

How to build and query a project workspace

Amir started this system when he took over an AI agents initiative at monday.com. With a thousand different internal opinions on what an "agent" should be, he looked for an unbiased external signal in public Reddit discussions. While candid, he knew these conversations were not a representative sample of monday.com customers or a substitute for direct research.

1. Start by uploading foundational context

Amir begins each new workspace by uploading documents that explain the product, team goals, and current initiative. This initial context makes later answers specific to the project.

Start with approved documents: Amir used kickoff decks, PRDs, internal product documentation, and public monday.com pages saved as PDFs. Before uploading any company material to an AI service, be sure to follow your organization’s data-classification, retention, and vendor-use policies.

Convert useful context into a readable format: Web pages and slides can often be saved as PDFs or text using a simple print-to-PDF command. The important constraint is permission, not technical convertibility. Only add material you are authorized to use in that workspace.

2. Use an AI assistant to write a data scraper

To collect public Reddit discussions about AI agents and monday.com, Amir used Claude to help him create and run a Python script. Instead of writing it from scratch, he started with a broad goal and let the AI guide him toward a workable solution.

Define the Problem Broadly: Start by telling the AI your big-picture goal.

Find public discussions about monday.com and AI across sites such as Reddit, X, and LinkedIn, and help me design an automated collection process.

At the time of recording, Claude steered Amir toward Reddit because its developer access was more workable for his use case. Platform APIs and terms change, so anyone repeating this workflow should use current official access methods and respect rate limits and content-use rules.

Ask for explicit setup instructions: Amir requested a beginner-level walkthrough, asking for instructions as if he were five years old. This helped him inspect each step and catch any missing configuration details.

Claude generated a Python script that used Reddit developer credentials. Credentials should always be stored in environment variables or a secret manager, never in a prompt, source file, screenshot, or repository.

The script looks for discussions related to Monday and AI or other specific keywords Amir provided.

The resulting CSV file contained 34,000 rows of public conversations.

3. Analyze the scraped data to find patterns

With the raw data in hand, Amir asks Claude to quantify recurring topics and cite the source rows. The product second-brain workflow covers collection, analysis, verification, and reuse without repeating the implementation twice.

Upload the Dataset: Give the AI the scraped .csv file.

Prompt for analysis: Amir requested a table that included each topic’s frequency, its percentage of the dataset, its relative weight, and representative discussion points.

Summarize the conversations in a table. Include each topic’s frequency, percentage of the dataset, relative weight, and representative discussion points. Cite the source rows so I can verify the findings.

Verify the analysis: Amir always asks for supporting excerpts or source references, then checks them against the original data. The citations help with the review but do not guarantee that the sample or the model's interpretation is correct.

Add the reviewed findings to the project workspace: Once Amir checks the analysis, he adds the summary to the initiative’s context. It remains one research input among many, alongside direct customer feedback and product data.

4. Configure the workspace to be a candid thought partner

Amir also gives the workspace instructions for how to respond. He hates when an AI is "super supportive" of every idea, so he configures it to behave like a product colleague who challenges weak reasoning rather than agreeing by default.

Define the role: Ask for product-management, strategy, and product-sense perspectives, then specify the types of decisions the model should help examine.

Ask for candor: Amir explicitly tells the model to question his ideas and explain its objections, which helps him pressure-test his own thinking.

Give me candid feedback. Challenge weak ideas and explain why you disagree instead of supporting every suggestion by default.

Amir uses the workspace to draft PRD outlines and answer questions about releases. The output is a working draft grounded in the uploaded context, not an authoritative record of product decisions.

Building a custom GPT to coach you on concise writing

After repeatedly hearing that his Slack writing was thoughtful but too long, Amir built a custom GPT that edits his messages for clarity and concision while trying to preserve his natural voice.

Step-by-step: Building your AI writing coach

1. Gather expert guidelines

Instead of just telling the GPT to "be concise," Amir supplied it with specific writing principles and examples from experts he trusted.

Use material you are allowed to provide: Amir referenced communication guidance from a Lenny’s Newsletter post by Wes Kao, as well as classic writing books. A reusable version of this workflow can use your own notes, licensed material, and short examples rather than uploading copyrighted works without permission.

2. Define the GPT's instructions and persona

The quality of this workflow comes from the specificity of the instructions. Amir tells the GPT exactly what good output looks like, what tone to preserve, and what stylistic habits to avoid, such as overusing dashes or bullet points.

• Core Function: "This GPT rewrites Slack messages to be more concise, to be more clear, to be readable."
• Preserve the writer’s voice: The instruction asks the GPT to keep the user’s natural voice rather than flatten every message into the same generic, AI-generated tone.
• Specific Style Directives: Amir also included negative constraints, like "avoid too much dashes, don't give so much bullet points," to fine-tune the writing style.

3. Daily application: The instant coaching loop

The workflow itself is simple. Amir drafts a message, pastes it into the custom GPT, reviews the rewrite, and sends the version that feels clearer and tighter. Over time, the repetition becomes its own form of writing practice.

Review and learn: Comparing the suggestion with the original can reveal repeated habits, but Amir still makes the final call on which version to send.

Amir reports that colleagues respond more often to his Slack messages, though the workflow does not establish causation. The custom writing-GPT workflow shows how he turns specific style guidance into a repeatable edit-and-review loop.

Practicing for product interviews with an AI mock interviewer

Amir uses GPT Voice Mode for live mock interviews and candid feedback. It gives him an always-available practice loop, while human practice partners remain essential for realistic judgment and interpersonal feedback.

Step-by-step: Mastering interview prep with AI voice mode

Amir had tried practicing for interviews by watching videos, talking to a mirror, and running sessions with friends. Voice Mode added a responsive interviewer that could ask relevant follow-up questions and adapt within the conversation.

1. Configure an interview coach GPT

You can just start talking to the AI, but creating a custom GPT gives you a more structured and informed interview practice. Amir made one just for this.

Define Interview Structure: Load the GPT with information about different types of interviews (like product sense or product execution) and how they usually flow. For example, a product sense interview often starts with clarifying the mission, then identifying users and pain points, and finally designing a solution.

For a product-sense interview, ask me to clarify the mission and product goal, identify target users and their pain points, and then design a solution. Use separate criteria for product-execution questions.

Add an interview framework you trust: To give the GPT context, Amir uploaded his own product-manager interview-prep document and posts from a product coach. Always use material you have permission to provide.

Provide only necessary personal context: A résumé and job description can help tailor questions, but be sure to remove sensitive details and consider whether the workspace’s retention settings are appropriate before uploading them.

2. Set the ground rules for the AI interviewer

Amir tells the interviewer not to lead him toward an answer or soften its feedback. These constraints keep the practice focused on his own reasoning rather than on the model’s hints.

Interview me for an AI product manager role at monday.com using a product-sense question.

Require independent answers and candid feedback: Ask the interviewer not to provide hints during the exercise and to evaluate the response only after the session is complete.

Do not guide me or give me hints. Let me choose the direction of the interview. When we finish, give me candid feedback on what worked and what I should improve.

3. Engage in a live mock interview

In Voice Mode, the GPT asks role-specific, product-sense questions and responds to Amir’s answers in real time, simulating the back-and-forth of a real interview.

Example AI Question: "Imagine you are tasked with improving the user experience for one of Monday.com's AI powered features. How would you go about identifying areas of improvement and what steps would you take to implement those improvements?"

4. Leverage continuous feedback

Across repeated sessions, Amir used the feedback to focus on weaker areas, such as user segmentation. Whether context carries between sessions depends on the product’s current memory and workspace settings.

Track progress against a rubric: To ensure consistent improvement, save the evaluation criteria and your own notes so progress doesn't depend only on the model’s memory or praise.

Amir says the practice made him feel "super prepared" before interviews. The Voice Mode interview workflow includes the role setup, follow-up questions, critique, and replay loop he used for rehearsal.

The verdict: What's worth copying from Amir's AI setup

The project workspace keeps initiative-specific context together. The writing GPT provides a fast editing pass. Voice Mode supplies repeatable interview practice. Each workflow reduces the effort required to resume a complex task or rehearse a high-stakes conversation.

None of the three systems removes the need for primary research, source checking, or human feedback. Their value is in speed and availability: Amir can generate a first analysis or a practice round, inspect it, and then decide what deserves further human work.

The most practical starting point is a single recurring task with a clear review step. Define which information the model is allowed to use, what evidence the output should cite, and which final decision remains with you.