Most PMs I know have hit this. You vibecode a working v0.1 in twenty minutes. v0.2 still works. By v0.3, the agent’s drifting — rewriting things it shouldn’t touch, breaking the styling, adding features you didn’t ask for. You ship a fix and it breaks two other things.

The agent isn’t the problem. It does what’s in front of it, in the order it’s in front of it. What it doesn’t do is push back, ask clarifying questions, or substitute good defaults the way a senior engineer would. Ambiguity in the spec becomes confident wrong choices in the code — and you find out at v0.3, when the consequence is now load-bearing.

You already know how to write a PRD. The skill transfers almost entirely. The delta is three additions:

1. Gating rules. Decision classes the agent must check with you before doing. e.g. “Stop and confirm before adding any third-party dependency.”
2. What-NOT-to-do. The negatives, drawn explicitly. PRDs are usually written in the positive; agents need the negative space named. e.g. “Do not use IndexedDB for the primary persistence layer.”
3. Escalation protocol. What the agent does when the brief doesn’t cover something. Including “do not pick a default and keep going.” e.g. “Pause. Write the ambiguity to QUESTIONS.md. Propose two options. Wait.”

PRD + three additions. That’s the whole framework. It’s the difference between “falls apart at v0.3” and “ships v1, v1.1, v1.2 cleanly.”

I’m running a free 90-minute live workshop on this. Bring a small idea — something you’d use yourself, something rough. We’ll walk it through the method end to end. You leave with a handoff doc that actually works and a clear read on whether to take it further.

Most analysts, consultants, marketers, ops folks I know use AI for the small stuff — polishing emails, summarising PDFs, brainstorming bullets. Useful, but a small upgrade on the old workflow.

The same tool that’s polishing your emails could be producing your entire weekly deck. Or running the Excel analysis you spend Wednesday afternoon on. Or making the interactive HTML explainer the team’s been asking for. Or generating the PowerPoint for next quarter’s board pitch. Not drafting parts of these. Producing the whole thing.

The shift isn’t no-AI to some-AI. It’s AI-as-thesaurus to AI-as-colleague. Same tool, different operating mode.

Three moves do the work:

1. Describe the deliverable, not the task. Not “summarise this” but “produce a one-page executive briefing, three sections, no jargon, bullets under each heading.” Pulls the AI up to artifacts, not subtasks.
2. Hand over the inputs the way you’d hand them to a colleague. Data files. Last week’s version. What’s important this week. The context you’d verbally explain — “the dip in metric three is the holiday.”
3. Tell it what good looks like. Your voice. What your readers won’t accept below. The format that’s worked before. The thing you know implicitly, made explicit.

Three moves, same shape whether the deliverable is a slide deck, a spreadsheet model, an interactive web explainer, or a position paper. The deliverable changes. The method doesn’t.

I’m running a free 90-minute live workshop on this. Bring something real — a deck you’ve been putting off, an analysis you don’t have time for, an explainer the team needs. We’ll walk it through end to end. You leave with the thing produced and the method to do it again.