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.

Grill UX

1. Open with a loose prompt asking the agent to grill the design — many questions, contradiction pushback expected. The grill happens in this session.
2. Agent asks many questions (what and why: purpose, empty states, build-vs-buy, security/XSS) while exploring the codebase in the background (non-blocking).
3. Answer a whole round in one dictation/paragraph, not one-liners. If 12 questions came back and you answered one, you are not done grilling — finish the frontier before execute.
4. Expect contradiction pushback — pressure is a feature.
5. Stop when the design frontier is empty — operationally, when every line below is YES:

• purpose statable in one sentence; scope edges named (explicitly what's OUT)
• empty/error states decided; build-vs-buy answered
• data flow sketched (entry → storage → UI); verification method chosen
• a full question round from the agent surfaces nothing new

Anti-patterns

• Plan mode as a substitute for grilling.
• Clear (or new session) before execute — throws away the grilled primary context; defeats the loop.
• Starting implementation while open grill branches remain.
• Treating the mini-spec document as the valuable artifact.

Execute

Stay in the same session so grilled shared understanding stays primary. Smallest vertical slice that proves the decisions. Verify with real checks — hallucination (fluent wrongness) is the default failure mode; never trust an unverified claim about your own codebase.

Model realities (don't fight them)

• Non-determinism — repeats differ, even at temp≈0 under parallel compute. Forcing max-likelihood often lowers judged quality (the likelihood trap). Control process (verify, small slices), not vibes.
• Statelessness — nothing carries between sessions unless written down: notes, skills, tracker specs, temp handoffs.

Clear (named loop default)

After the feature is done and committed, Clear before an unrelated next task. For non-default boundaries (QA, cross-agent, AFK review), use the phase-boundary tree in handoff / compact.

If grilling reveals a multi-session feature

This loop is single-session. The mini-spec graduates: hand it to to-spec as the seed of a destination spec + vertical tickets. The mini-spec stays a confirmation note — it never becomes a living repo doc.

Output

• Decisions that still bind
• What landed + verify
• Next phase + continue | clear | handoff | subagent | compact (tree order)

Related

to-spec · handoff · compact