You have a comprehensive spec. You don't.

I keep seeing product hand a markdown spec to engineering and call it a plan.

Product owner does the normal Scrum thing.

They specify what they want.

They don’t get to decide how it’s built. Fine. That’s the working agreement.

Now they have an agent look at the code, work through a few possible outcomes, and spit out a markdown file.

A PRD, basically. Looks thorough. Looks like someone spent a week on it.

That first agent actually had the good stuff. The weird code paths. The conversation. Why one option felt wrong. What nobody decided yet.

Then it writes the document like a product owner is supposed to write a document. Dumbed down on purpose. Vague on purpose. Pretending to be a human artifact.

Engineering takes that file and feeds it to their coding agent.

And of course the code is never quite right the first time.

Same missing gap planning always had, except now it looks like you did the planning. You have a comprehensive spec. You don’t. You have the worst kind of compaction of a conversation.

I keep thinking about Tom Smykowski in Office Space. “I take the specifications from the customers and I bring them down to the software engineers.” If you’re walking a cleaned-up markdown file from one agent to another, that’s you.

You’re a “meat harness.”

I don’t think you dump the whole transcript into the next prompt. That’s too messy.

But the next agent needs the source context and the decisions. Not the pretty summary.

I suspect you’d do better work if product and engineering were in the room at the same time, looking at the same thing, instead of handing an artifact over the wall. That’s always been a little true.

Agents just make the cost of the wall more obvious.

(This is probably going to be great for product engineers. That’s another post.)

If you’re treating the AI spec as the source of truth because it looks more complete than a human one, I wouldn’t.

It just looks the part.