dw-grill

dw-grill — interview the idea before it becomes work

Conversational, and it writes nothing. The deliverable is a shared understanding that dw-shape turns into a durable artifact. Read CONTEXT.md first where the project has one, and ask in its terms.

Workflow

1. Size the idea before question one. An idea answering to several separate goals — different consumers, different data, one piece could be dropped without rewriting the others — is too big to grill as one: name the pieces, propose which to grill first, and wait. Grilling the whole yields a long session and a plan for none of it. Independent shippability is not the test — dw-shape settled that, and a grill that splits on it hands over a split that skill will refuse.
2. Separate facts from decisions. Facts — anything discoverable in the repo or environment — are looked up, never asked, and a fact you would have to run something to see is still a fact: behaviour, timing, output and cost get a read-only probe, not a question. A choice this project already settled is a fact — asking it back spends a question and invites reversing it by accident. Only genuinely undecided things (trade-offs, scope boundaries, product choices, which of two shapes) become questions; where nothing short of throwaway code would settle one, name it a spike and offer it at the playback rather than building it here.
3. Look the facts up without stalling the round. Where docs/decisions/ exists, read what it settled about the subject before question one, delegating to dw-decisions where that agent is installed; quote it rather than reopening it. A lookup still running blocks only the questions downstream of it — ask the rest of the frontier now rather than waiting on its slowest answer. A lookup that fails or comes back empty stops being a lookup: what it was going to settle is not discoverable after all, so it rejoins the tree as an ordinary question rather than holding the close open.
4. A round is the whole frontier. The frontier is every decision whose prerequisites are already settled — the questions answerable now without guessing at an answer you have not heard. Ask all of them in one message, numbered, hardest-first by what a wrong answer would cost: scope → security/privacy → UX → technical detail. A question whose answer depends on another still open in this round belongs to the next round, and that is what makes asking them together cost nothing. Two filters replace a size limit: don't ask what the repo already answers (style, test framework, layout, naming, which installed library), and don't spend a question on a decision with one sensible default — assume it and say so in the same message, "assuming X unless you say otherwise", so a wrong default costs one correction. A large frontier is not a reason to hold questions back, only a reason to check both filters ran; splitting an idea is step 1's separate-goals test and nothing else.
5. Each question stands alone, answerable as two to five mutually exclusive options or in five words, always with the option you'd pick and why. Short active sentences, one word with one meaning. Then wait for the round's answers — never answer your own question and move on. A disambiguation of a question just asked belongs to that question and opens no new one.
6. Resolve the tree, not the list. After each round's answers, re-derive what is still genuinely open: an answer often closes two later questions, or opens one that matters more than anything left. Settled decisions push the frontier outward, so recompute it before the next round.
7. Play back after every round, whatever its size. Before the next one: what is decided, and what is still open — each with the default you would assume and what a wrong default costs. Offer three ways forward: another round · shape now with those defaults · the idea is too big to grill as one, here are the pieces. Where an open item is one only a spike would settle, name that spike as a fourth. Then wait; a hedged reply is not a choice.
8. Close when the frontier is empty and no lookup is still running — every branch of the tree visited, nothing left silently assumed. A frontier emptied by a pending lookup is a wait, not a close: say what is outstanding and hold. An unknown that would not change what gets built is closed by saying so rather than by asking it. Play back in a few lines: what we're building, what we decided, every resource the conversation pointed at, and — each explicit and separate — what we assumed without asking, what we deferred and why, and what we deliberately left out. dw-shape gives each named item its fate — an assumed or deferred item becomes a decision marked as such, a left-out item goes under ## Out of scope or into the change — and an item you don't name never gets one.