When to stop talking to the AI, and start building

You set up a workflow with Claude and have been using it for a few weeks. It works well.

But when you run it today in a new session, the results don’t feel right. Or you spent several minutes explaining the context that you thought you’d already stated before.

Sound familiar?

This is prompt drift. And it’s easy to miss because it used to work so well - and drift can happen gradually.

When you’re figuring out a new workflow, prompting is a good place to start. It’s easy, and you’ll see results quickly. You can reshape the approach mid-conversation, test variations, and get to something useful without building anything.

The prompts are effectively a prototype. You’re validating the workflow before investing in it.

The problem is when you stay in this phase too long.

Here are the signs that prompting has stopped being the right tool:

• You’re running the same prompt more than once a week

• You’re re-explaining context at the start of every session

• The results drift - close enough one day, noticeably off the next

• Someone else couldn’t replicate your process without you walking them through it

If the workflow lives inside your head and a chat window, it’s not really a workflow. It’s a workaround.

The obvious AI-era answer would be to automate it - schedule the prompt, chain it with something, let the AI run it on a timer.

But you’ll still see the same issues with “prompt drift”.

Instead: stop prompting it and build it into a product.

When a process is genuinely repeatable - same inputs, same logic, same expected output every time - it doesn’t need a conversation. Code it. Script it. Turn it into a proper tool.

The AI conversation is great for working out what the process should be. But it’s a poor substitute for the process itself.

Prompts are non-deterministic by design. That’s what makes them powerful for exploration. It’s also what makes them a fragile foundation for anything you’re relying on week after week.

As PMs, we need to spot the difference between a prototype and a product.

Prompting an AI workflow is the prototype phase. A prototype helps you learn - and this can take you a long way.

A product is what you ship when you know enough to commit. At some point, you need to graduate to a product.

Product Notes is free and lands every week. Subscribe to get it in your inbox.