machine unlearning

(Edited with Claude)

My first attempts using GenAI chatbots to build apps gave results that were flashy but lacked substance. A few weeks ago, I got frustrated using iMessage to coordinate a spontaneous hangout with friends. What was supposed to be an easy trip to our local bar turned into hours of back-and-forth texts. By the time we settled on a time, my internal battery was depleted.

Curious, I did what most people are probably doing—I opened a few chatbots and entered the same prompt:

I want to build an app that helps me schedule casual and spontaneous hangouts with friends.

It was magical watching applications build in real-time, but the solutions were more complicated than what I actually wanted to use. Claude and Lovable immediately started building. Cursor was different—it prompted me with key questions to scope the idea down.

Claude

Lovable

Cursor

This got me thinking: is there a product equivalent to prompt engineering? A set of techniques to guide genAI chatbots into producing answers that align with product thinking? We know the LLMs matter, but prompt engineering shows that how you interact with a chatbot is just as important.

---

A Framework for Product-Led LLM Development

Over the past few weeks, I've been experimenting with simple product development using gene chatbots. I'm sure my techniques will evolve—soon I might even have my own agent based on my individual style of product development. But here's a starting point for using chatbots with product thinking.

Start with the User Problem

Like any product framework, begin by identifying the user problem you're trying to solve. The advantage of feeding this into a chatbot is turning it into a thinking partner. You're not solutioning yet—use the bot to think big and move quickly through possibilities.

Identify Your User Segment

Just because it can code before your eyes doesn't mean the solution will be valuable. You still need to understand who you're building for. Once the chatbot knows your user problem, prompt it to ideate on potential user segments and select the one that best fits.

Define User Success

Now that you have your problem and user segment, identify what success looks like: What actions should users be able to complete? How should they feel when using your application? This should inform the solution, not the other way around.

Set Constraints and Ideate Solutions

With user success defined, start ideating on solutions with constraints: platform (web, mobile, script), level of polish (MVP/MLP), tech stack preferences, and core features. The more you define, the more tailored the chatbot’s recommendations become.

Select a Solution and Ask for Architecture Options

As the PM/Engineer, you still need to use your “taste” to identify if the architecture of a solution will work in your context. For example, chatbots aren't great at building scalable applications yet. Try the common prompt engineering technique of asking for alternative architectures or question whether the recommended architecture is scalable.

Why This Matters

With these techniques, you can use product thinking to guide genAI chatbots toward viable solutions without building anything first. This criteria can become a detailed PRD (you can even use the bot to review it), or you can prompt it to start building the solution and architecture you’ve identified immediately.

The result? Higher control over the solution and something that's actually valuable to potential users, not just impressive to watch being built.