Every effort starts with a grill: the agent relentlessly questions the idea before a single line of code exists. Not a friendly brainstorm — an interrogation. What problem does this actually solve? What breaks if we don't build it? What's the cheapest version that would teach us the same thing?
The direction of the questioning is the point. The agent grills the human, because the human holds the context that isn't written down anywhere yet — and the grill is where that context gets extracted, challenged, and pinned to the page.
How it runs
This is the one step that cannot run agent-solo. Everything downstream — the spec, the tickets, the implementation — inherits whatever understanding the grill produced. An agent answering its own questions here would just launder assumptions into requirements. So the grill resolves only in live exchange: the agent asks, the human answers, and the step ends when the human explicitly confirms the shared understanding — never when the agent decides it has heard enough.
Terms and decisions are recorded at the moment they crystallize, not reconstructed afterwards. When the grill coins a word or settles a trade-off, it lands in the project glossary or a decision record immediately — as a side effect of the conversation, not a documentation chore at the end.
The exit condition has a second half: the conversation continues unbroken into writing the spec. Summarizing a grill and replaying it to a fresh session loses exactly the nuance the grill existed to surface — so grill, spec, and tickets happen in one continuous window.
What gets settled here
Several decisions belong to the grill, precisely because they're cheapest to change while everything is still words:
- Technology choice, governed by a reuse ladder: an existing pattern already in the codebase, then a battle-tested library, then a proven external implementation worth porting, and only last, net-new code. Stepping down a rung has to be argued, not assumed. Choices that are hard to reverse and genuinely surprising get a decision record.
- Whether to pause for research. The tell is a sentence like "I think X probably supports this" — a decision resting on an out-of-repo fact neither party can state with confidence. That fact gets verified against primary sources, and the answer is read back into the grill.
- Whether words have stalled. If a decision keeps circling after a couple of rounds — how something should behave, or how it should look — the answer is a throwaway prototype to react to, not another round of adjectives.
- Whether this fits in one sitting. At the end of the grill comes a simple test: could we write the spec and its tickets right now? What forces a longer mapping effort is fog — open decisions nobody can answer yet — not size.
Built on
This step is built on Matt Pocock's grilling skill, extended to capture glossary terms and decision records as the conversation runs.