Steering what gets built

The gate

Before your agent authors or rewrites any document in the chain, it runs its elicitation playbook: a structured round of questions put to you, whose answers become the document content. This is a hard rule of the method, not a courtesy. If a fact was not given to the agent, it is not allowed to invent it; a chain written in one shot without your input is a method violation, not a deliverable. That is the default, not a limit on what the agent can do for you. Ask it to help invent the product, or to fill a gap with its best guess, and it will; the gate is the safety net that stops invention happening without you asking for it. The gate opens at the start of a project and again whenever new capability enters the tree.

What it will ask you

The questions come in plain language, a few at a time, and they cluster into four areas. The problem and who has it, which frames the product. What the product must do and for whom, which becomes the requirements and user stories. What done means, concretely enough to test, which becomes the acceptance criteria, and this is where the agent will push you on edge cases you had not considered: the empty result, the duplicate entry, the second user. And where the thing will run, because the deploy target constrains the technology choice, so the method insists that question is settled, or explicitly parked, before a stack is committed.

If the round feels short, that is deliberate, though you can pause at any point and dig into as much detail as you need. The playbook tells the agent when to stop asking and start writing, and you review everything it wrote at the docs-review gate before any build starts.

"I do not know" is an answer

You do not need answers to all of it, and pretending is worse than deferring. Tell the agent you do not know, or that you have not decided. It records the deferral as an architecture decision on the tree and holds back the live half of whatever was deferred. Deferring hosting, for instance, does not stall the project: the agent builds you a working local preview instead, an app you can run on your own machine, while the deferred decision sits in the tree, visible, waiting for the day you settle it. What the agent must not do is press you into standing up accounts and billing for a thing you have not committed to.

Changing tack

Elicitation is a conversation, not a form. If the questions are going somewhere you do not care about, say so mid-round:

AIPark the hosting question, I have not decided. Search matters more than capture; ask me about search first, and keep the first version small.

The agent adjusts and the tree follows your priorities, not its assumptions. You can also reopen the gate at any time: "I have changed my mind about who this is for, re-run the questions for that part" is a legitimate ask, and the method treats the resulting edits like any other change, validated and traceable.

Reading the questions for yourself

The playbook the agent works from ships inside the package, word for word. Ask your agent to show it to you if you want to see what is coming, or read the elicitation playbook here.