𝗔𝗜 𝗣𝗿𝗼𝗱𝘂𝗰𝘁 𝗠𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁 𝗶𝘀 𝗼𝘃𝗲𝗿𝗵𝘆𝗽𝗲𝗱. That is what I told myself when I first started working on AI products. Turns out, I had no idea what an AI PM really did. After spending years watching world-class AI PMs, building AI products at scale, making 100s of bad decisions, I have a much better definition of the role. But, sadly, even today, most PMs trying to work on AI are in the same boat: confused and clueless about "what does an AI PM actually do." Here's the simplest mental model to think of an AI PM's role: An AI PM is responsible for finding answers to these 7 questions. 1. What problem should we solve to maximize impact? 2. Does this need AI? 3. Do we have the right data? 4. How do we turn data into something useful? 5. How will users experience it? 6. How do we know it works before launch? 7. How do we keep making it better? Let's understand in detail: 𝗪𝗵𝗮𝘁 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 𝘀𝗵𝗼𝘂𝗹𝗱 𝘄𝗲 𝘀𝗼𝗹𝘃𝗲 The problem must be specific, validated with real users, and solution-agnostic. If you get this wrong, the model does not matter. 𝗗𝗼𝗲𝘀 𝘁𝗵𝗶𝘀 𝗿𝗲𝗮𝗹𝗹𝘆 𝗻𝗲𝗲𝗱 𝗔𝗜 This is the most important question an AI PM asks. And the answer is usually no. Saying no to AI when the situation does not call for it is not a failure. It is the job. 𝗗𝗼 𝘄𝗲 𝗵𝗮𝘃𝗲 𝘁𝗵𝗲 𝗿𝗶𝗴𝗵𝘁 𝗱𝗮𝘁𝗮 "We have data" is not a strategy. An explicit data plan that includes what data we need, what we have, and what is missing is the right strategy. AI is only as good as the data behind it. 𝗛𝗼𝘄 𝗱𝗼 𝘄𝗲 𝘁𝘂𝗿𝗻 𝘁𝗵𝗲 𝗱𝗮𝘁𝗮 𝗶𝗻𝘁𝗼 𝘀𝗼𝗺𝗲𝘁𝗵𝗶𝗻𝗴 𝘂𝘀𝗲𝗳𝘂𝗹 Simple prompt, ML model, RAG, or agents. Each has a different use case, cost profile, and failure modes. The PM who skips this hands those decisions to engineering. 𝗛𝗼𝘄 𝘄𝗶𝗹𝗹 𝘂𝘀𝗲𝗿𝘀 𝗲𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲 𝗶𝘁 Users need trust, control, and recovery. Design for when the AI fails, not only for when it works. 𝗛𝗼𝘄 𝗱𝗼 𝘄𝗲 𝗸𝗻𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀 𝗯𝗲𝗳𝗼𝗿𝗲 𝗹𝗮𝘂𝗻𝗰𝗵 There is no binary pass/fail. Build an eval framework. Define good, bad, and edge cases. Ship only when the product clears your threshold. 𝗛𝗼𝘄 𝗱𝗼 𝘄𝗲 𝗺𝗮𝗸𝗲 𝗶𝘁 𝗯𝗲𝘁𝘁𝗲𝗿 𝗮𝗳𝘁𝗲𝗿 𝗹𝗮𝘂𝗻𝗰𝗵 AI products degrade in production if you stop watching them. Sample live conversations. Add new failure modes to your test set. Never stop monitoring. -- Want more details? I just published a full document covering everything I know about AI Product Management--the role, mental models, the mistakes I made and much more. Linked in the comments. Free. No email required, no paywall. I will keep it updated as the field evolves.

Most people think Claude code only needs good prompting. Unfortunately, that is not true. The real unlock is giving it a structure that is coherent and can last from start to end. If there is no structure, Claude code is just another LLM If there is proper structure, Claude code is the smartest engineer on the planet. Here are the 6 files you should include for every AI-assisted build: #𝟭 𝗣𝗥𝗗 This is your master plan. What you're building, who it's for, what good output looks like. Created once. Iterated rarely, only if absolutely necessary. The LLM refers to it after very major milestone for context, so it doesn't guess your intent. #𝟮 𝗣𝗹𝗮𝗻𝗻𝗶𝗻𝗴 This is the product's blueprint. User flows, screens, navigation, system diagrams. It defines all the details and the structure so the LLM doesn't invent pages you never asked for. #𝟯 𝗧𝗮𝘀𝗸𝘀 This is the to-do list. Milestones and tasks with checkboxes. The LLM reads this to know what's done and what to build next. Updated every session. This helps the LLM know exactly what is done and what is pending, so you don't have to remind it. #𝟰 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 (𝗜 𝘂𝘀𝗲 𝗖𝗹𝗮𝘂𝗱𝗲[.]𝗺𝗱 𝗳𝗼𝗿 𝘁𝗵𝗶𝘀) This your product's soul. Limit to 300-500 words. Include one-paragraph product summary, tech stack, design preferences, "never do" list. The LLM reads it before every action. When context gets lost in a long session, this file is what keeps it on track. It ensures that every action is taken in the right direction. #𝟱 𝗗𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝘀 Your learning log. It incldues every critical decision, every tradeoff, every "we went with X because Y." It logs what you decided to do and why, so you can refer back reflect / learn from your decisions and get better every time. This is the only file that's 𝗙𝗢𝗥 𝗬𝗢𝗨, not the LLM. This is what makes you a better AI builder. #𝟲 𝗠𝗲𝗺𝗼𝗿𝘆 This your product's long-term memory. It includes critical context that must survive across sessions. Bugs that were fixed, patterns that worked, things that broke. Claude Code reads this so it never repeats a mistake or forgets a lesson you already paid for. Knowledge tells it who your product is. Memory tells it what your product has been through. -- This is the shift you need to make. Prompting is necessary. But it is not enough. Aim to create a structure that AI understands Once you do that, Claude automatically becomes the best collaborater and executes like a high performing engineering team. Reply "structure" and I will create a detailed post on each of the 6 files with real examples and snippets.