grill

The most expensive failure in development is building well what wasn't wanted. And it is almost never detected at the start: it is detected when the result is shown.

A brief arrives with gaps. The agent fills them in alone and in silence, because filling them is faster than asking. Every gap filled by eye is a bet, and a single one going wrong throws away the afternoon's work.

This skill turns those gaps into questions before they cost anything.

The tree and its frontier

A plan is a tree of decisions: each one opens the ones that hang from it. You can't ask about the PDF format if you don't yet know whether there will be a PDF.

The frontier is the decisions whose prerequisites are already resolved: what can be asked now without guessing answers you haven't heard yet.

Work in rounds:

1. Compute the frontier.
2. Ask the whole frontier at once.
3. Wait for the answers. Don't move on without them.
4. The answers push the frontier outward and unlock what depended on them. Go back to step 1.

A question whose answer depends on another that is still open in this round belongs to the next round. Slipping it in forces the user to answer twice or to invent a condition.

It ends when the frontier is empty: all branches visited, nothing assumed in silence.

How a round is phrased

Each question is numbered, with a title, and with the answer you recommend. The recommendation is not optional: it forces the agent to commit and lets the user say "yes" instead of writing a reply.

❓ **Q1 — <short title>**

<the body of the question: context, options, what changes depending on the
answer. It can run to several paragraphs>

➡️ <what you recommend and why, in one or two sentences>

---

❓ **Q2 — <short title>**

…

If the agent has a closed-options control, use it when the alternatives are few and mutually exclusive, with the recommended one first. For everything else, the format above.

The rule that holds it all up

The agent looks up facts. The user makes decisions.

If a frontier question needs a fact about the environment (which version is installed, whether that endpoint exists, how many rows that table has), that is not asked: it is looked at. Asking it spends the user's turn on something the agent resolves with a command, and on top of that gets an answer from memory instead of a fact.

Don't block on it. An exploration in progress is an unresolved prerequisite, so only the questions that hang from it wait; the rest of the frontier is asked right away.

When to stop

This skill is expensive in user turns, so it has a ceiling:

• Two or three rounds for a normal change. If you are on the fifth, either the brief is huge and has to be split, or you are asking things you should be looking at.
• No question whose answer doesn't change what you are going to do. If you build the same thing with both possible answers, the question is superfluous.
• No confirming what was already said. If it is in the brief, it is decided.

Afterwards

The result of an interrogation is an agreement, and an agreement that only lives in the conversation is lost when the session closes. Move it to where it belongs:

• /explain-plan if the proposal has to be presented before building it.
• /adr if along the way an alternative was discarded for a reason that isn't obvious.
• crux task if the work will last more than one session.

Don't start building until the user confirms you are in agreement.

The mechanism of the tree, the frontier and the rounds is adapted from the grilling skill of mattpocock/skills, MIT license.