grill-execute-clear
grill-execute-clear
Loop for agent-assisted feature work: grill until the design frontier is empty → execute in the same session → clear (or compact/handoff/subagent) at the phase boundary.
Same-session execute after grilling. Never clear-before-execute.
When
Non-trivial feature where explore→implement with no alignment wastes the smart zone. Prefer this over raw plan mode.
Why plan mode fails (Brooks punch)
Frederick Brooks (The Design of Design): a design concept is not an asset. It's the shared picture in the room while humans align — not a document you can bank. Plan mode skips that alignment and rushes a plan that is the implementation (agents are sycophantic). Specificity of an early plan is often the bug. Grilling forces decisions first; any written mini-spec is confirmation, not discovery. You usually do not need to read the generated mini-spec if you answered the grill honestly — the shared understanding already lives in the primary session.