Grilling

Interview mechanics. A sharp question up front is cheaper than a wrong document; during an interview, asking is the job — the usual "only ask when you cannot proceed safely" restraint does not apply.

A flow skill always owns the document

This skill owns the question tools and nothing else — it has no artifact contract, so fired on its own it produces well-formed questions with nowhere to put the answers. The document belongs to a flow skill: start for requirements from an idea, amend for a scoped change to them, design for the design. Each loads this one for the mechanics.

Reached directly, hand it back: name the flow that covers what the user wants, and let them fire it.

The tools

Each form ends your turn; the user's answers arrive as the next message.

• ask_questions — a form of several INDEPENDENT questions answered

together. Up to 8 per form: a screenful, not a budget. What does not fit goes in the next round.

• ask_question — one question, when the next question depends on this

answer.

Writing questions

• Ask only questions whose answers change the artifact. Everything else is

either already answered (the brief, an earlier answer, an org default) or yours to assume.

• Give 0–5 concrete options and mark exactly one recommended: true.

Add a description only where the label alone is ambiguous, and keep it to ONE short clause — the form is generated in full before the user sees any of it, so every extra sentence across every option is time they spend waiting on a blank screen. multiSelect: true only when several options genuinely co-apply.

• The form always offers free text, so pass an empty options list when the

answer must be typed — never invent placeholder options like "Other".

• Options are a starting point, not a cage: never smuggle in a constraint the

user didn't state; ask instead.

The answers are the authoritative brief — treat them as decisions, not suggestions.

How many rounds

As many as the work needs — no cap on questions, forms, rounds or sessions, for requirements and for design alike. Revisiting something answered earlier is legitimate when what you have written since changed what the answer has to carry; say what changed when you do.

What keeps that from being an interrogation is the order, not a budget:

• Reach the document early. Ask first what the artifact cannot be written

without, write it, then refine it in place. The calling skill says what its own document needs first.

• Every round after the first reacts to something concrete — a line the

document now carries, an answer that opened a new choice, two statements that contradict. A question that could have been asked before the document existed belongs in the first round. Without that, more rounds are strictly worse than the single form they replace.

• No denominator. State where the work stands by what the document says —

here is what I have; what is wrong with it? A convergence-based interview has no total, so any count you offer ("question 3 of 5", "one last round") is a fiction.

Ending the interview

• Converged: the remaining unknowns no longer change the artifact — stop

and generate.

• Recommended answers: the user says "just generate" / "skip", or takes the

form's own exit — stop asking, apply your recommended answer to every remaining decision, and tag each one assumed where it lands in the artifact. This ends the round, not the conversation: a flagged assumption is a decision the user can see and overturn, which is what makes it a deferral rather than a loss. A silent one is an invention.

• Deferred: the user says they know and will decide later — write "deferred"

against it and stop raising it. This is the one answer that does close a line of questioning for good.

• Headless: the turn states no interview is possible — ask nothing;

generate on stated assumptions, each marked assumed.

A judgment the agent can make is assumed and flagged. A fact only the user holds — a URL, a package, which vendor they have a contract with — is never assumed: it goes to the document's open questions, because an invented API URL does not fail at review, it fails at build.