The walk before the draft

I work through half a spec before I’ve typed a single word.

Not at my desk. On a walk, or on the treadmill before the day starts. Phone in my pocket, talking out loud to an AI.

It has nothing to do with multitasking. It’s about what happens when you can’t clean up your thinking before you say it out loud.

Why raw beats polished

At a keyboard, you edit as you go. You reach for the right word, restructure the sentence, delete the awkward part before anyone sees it. That instinct is useful for prose. It’s the wrong instinct for the first pass of a spec.

When you talk a feature out loud, none of that is available. No cursor, no formatting, no tidy template to hide behind. You just say what you actually mean. That messy, unedited version is the most honest input you’ll get all day. It’s the raw signal before your brain “helps” by polishing it into something that sounds more finished than it actually is.

The flip

Here’s the part that actually changes the output. Instead of asking the AI to write something, I have it ask me.

Three or four questions. The persona. The edges. What I’m sure of and what I’m only guessing. This is the Discovery step in the Prompt Blueprint framework I use for every AI-assisted PM artifact: the model interrogates the input before it drafts anything, instead of the other way around.

Flipping the question burden onto the AI is the whole trick. It surfaces the gaps I’d otherwise hit three sprints later in engineering, except now I hit them on a walk instead of in a stand-up.

By the time I’m back at my desk, the gaps are already on the table.

What actually comes out of it

I don’t come home with a finished spec. I come home with a clean list of what’s confirmed and what’s still a question.

In practice, this usually means a saved project with instructions tuned to push back rather than agree, so it acts like a sounding board instead of a yes-man. I talk through the problem, the goal, what I know, and what I don’t. I’ve used this for a handful of product features, and the pattern holds regardless of which one: talking it through out loud surfaces angles I hadn’t considered, and it leaves a record I can pick up later instead of trying to reconstruct my own thinking from memory.

Which tool depends more on logistics than principle. Some days it’s Claude, some days it’s ChatGPT, depending on which one I have credits left on. The habit matters more than the platform.

Naming the unknowns is the whole point. A flagged “I don’t know yet” beats a confident assumption that breaks the build three sprints later.

The takeaway

The keyboard is where you format the work. The walk is where you actually think it through.

Those are different jobs. Treating them as the same one is a big part of why so many specs read like they were written to fill a template instead of to solve a problem.

Where do you do your clearest thinking? If it’s not at a screen, I’d like to hear where.

If this way of running the loop is new to you, the Prompt Blueprint framework is the fuller method behind it: Context, Role, Discovery, Execution. Read it here →