Prompts are dead, the plan is the prompt

Prompts are dead. For two years the craft was prompt engineering: typing a better sentence into a box and hoping. That is over. The model is good enough that the sentence no longer matters; what matters is the plan behind it. The unit of work is now a plan, and the quality of the plan is the quality of the output. Vague plan, average AI slop. Precise plan, a system that runs.

Everyone has more ideas than they can ship. The constraint is never ideas. The constraint is the discipline to break ideas into things that can ship in a week, hold yourself accountable to that timeline, and resist the constant pull to expand scope mid-build.

This is the protocol that turns I have an idea into I shipped a working feature on Friday. Without it, the four agents chase shiny things, the project never finishes, and three months later you have a half-built mess instead of a working system.

Prompts are dead

For two years the craft was prompt engineering: a better sentence in the box, a clever persona, a magic phrase. That era is over. The model is good enough that the sentence is no longer where the leverage is. The leverage is in the plan behind the sentence: what exactly is being built, in what order, and how you will know it is done.

So you stop writing prompts and start writing plans. You do not even write the plan; you answer questions and the PM agent writes it. Vague plan, average AI slop. Precise plan, a system that runs every day. You direct, AI writes.

The five questions the interview asks

Every plan starts as a workflow decomposed. The PM agent asks these five, one at a time, about the business you already know. Done well, this turns a vague job into a shippable process. Done badly, you get a Zapier flow nobody understands.

• What is the work, in one sentence?
  If you cannot say it in one sentence, you have two products, not one. Split.
• What are the steps, three to eight of them?
  Fewer than three and it does not need a system. More than eight and it is two epics.
• Who does each step today?
  Be honest. You, a person, a tool, or no one. The no-one steps are usually the bottleneck, and the reason you are here.
• Which steps can AI do?
  On its own, as a draft for you to review, or never. The honest answer per step matters. Wishful answers compound into bad plans.
• What is the routing, written down?
  AI drafts versus AI decides is the distinction that matters most. Name it per step. From then on the plan, the agents and you all read the same document.

The hierarchy: Idea, Product, Epic, Task

The constraint on shipping is never ideas. The constraint is the discipline to break ideas into something a team can finish in a week. Four levels, in order.

• Idea
  Conversational, lives in a doc, revisited weekly. Most ideas die at this stage on purpose. The few that survive a week earn the right to be a product or an epic.
• Product
  A coherent body of work that has its own user, value proposition, and scope. "The Infinite Leverage website" is a product. "Lead capture system" is a product.
• Epic
  A multi-day chunk that ships a complete capability. Sized 3 to 7 days end-to-end. "Add Supabase contact form with admin view" is an epic. "Add a button" is not.
• Task
  Sized to a single session, 30 to 90 minutes. "Create contacts table in Supabase." "Wire form submit to insert API." Tasks roll up into epics.

Two rules that prevent 80% of mid-stream confusion

These rules are unfashionable. Project management tools want you to skip them. Skip them and you spend the next quarter recovering.

• Define the epic before you start the work
  Even a one-paragraph description: when this is shipped, a user can do X, the data flows Y, the UI has Z. Three sentences. Saves a week of scope creep.
• One epic at a time per person per agent
  Not two. Not three. One. You can have many epics open across your team, but each operator (human or agent) is on exactly one. Switching costs are the silent productivity killer.

The interview, then two documents

You do not write the plan. You answer questions, and the PM agent writes it. One question at a time, in plain English, about the business you already know. Then it produces two files, and the split between them is the whole trick.

• Working Files/product-plan.md, for you
  What we are building and why. The data model in sentences, not SQL. The epics, each a paragraph. A Definition of Done: a list you can check with your own eyes on the live site. This is the document you judge the build against.
• docs/plans/build-1-plan.md, for the AI
  Ordered steps, each with a visible check (a URL loads, a row appears, an email arrives). Under two pages. The developer agent executes this top to bottom and stops when a check fails.
• The Definition of Done
  Written before the build, not after. If you cannot say how you will know it is finished, you are not ready to start. QA tests every line of it (Protocol 16).

The PM agent holds the line

The PM agent's job is to enforce this discipline so you do not have to. It refuses to start epic 2 until epic 1 is closed. It refuses to break down epic 3 until epic 2 is started. It is the part of you that holds the line when you are tempted to chase shiny things.

A product plan in your words, a Definition of Done you can tick, and a build plan the AI can execute. Epics 2 to 6 intentionally left as one-liners. You have a plan you can actually build from, and a way to know when it is done.