jonmumm Auto Grill Me
A self-driving extension of /grill-me. Same interrogation, but the agent stops waiting on the human for each answer — it picks the most defensible option, records the choice, and rolls into the next question. The human's job shifts from "answer every question" to "interrupt only when something looks wrong."
When to invoke: in the middle of a /grill-me session, once the last several Q/As have been you saying "yes, obviously" or "go with your recommendation." That's the signal the agent has enough context to drive the rest itself.
When NOT to invoke: at the start of a grilling session, on a domain the agent doesn't know, or any time the alignment loop is still doing real work. Skipping that loop is the entire risk — see "Tradeoffs" below.
Behavior
When this skill is invoked, adopt the following directive for the remainder of the session:
Keep grilling. Walk the same design tree /grill-me would walk — dependencies between decisions, edge cases, hidden assumptions. Do not soften the questions to make auto-answers easier.
Answer each question yourself with the option you'd recommend, in one short paragraph: the choice, the one-line reason, and the assumption it depends on. Format:
Q: <the question you would have asked> A (auto): <choice> — <why> — assumes <assumption>Surface low-confidence answers explicitly. If your confidence in the auto-answer is low, or the question is load-bearing for the whole plan (architecture-defining, irreversible, or trades safety for speed), mark it:
⚠ low-confidence / load-bearing — interrupt me here if you disagreeDo not skip these. They are exactly the questions the human needs to see.
Run in short batches. Emit 3–5 Q/A pairs, then pause for one beat (single line:
— continuing, interrupt to redirect —) before the next batch. This gives the human a natural place to press Esc without losing context.End every auto-run with the /grill-me close-out question: "What did we assume that we didn't write down?" List the assumptions accumulated across the auto-answers, so the human can scan them in one place even if they didn't interrupt anywhere.
Tradeoffs (read before using)
This skill removes the human-in-the-loop friction that gives /grill-me its value. Matt's own original framing (the tweet this skill is adapted from) ends with "the more I think about it, the more I hate it" — because without the human, the agent converges on its own biases instead of the user's intent. The thread surfaced three real failure modes:
- Lost alignment. The point of /grill-me is shared understanding. Auto-answering trades that for speed.
- Missed non-obvious questions. A run of "yes, obviously" answers can lull both sides into auto-piloting past the one question that actually mattered.
- Hard to undo. If a bad auto-answer five questions back changed the design's direction, unwinding it is expensive.
Mitigations baked into the behavior above: explicit low-confidence flags, batched pauses, and a mandatory assumption-dump at the end. They reduce the risk; they don't eliminate it. If you find yourself interrupting more than ~1 in 5 auto-answers, stop — you're not actually aligned, and /grill-me's normal mode is the right tool.
Relationship to /grill-me
/auto-grill-me does not replace /grill-me. It is a mode you enter mid-session. The /grill-me directive (relentless interrogation of the design tree) still holds — auto-grill-me only changes who supplies the answers.
To exit auto-pilot, the user types anything other than "continue" / "keep going". Any substantive message — a correction, a new question, a "wait, go back to X" — drops you back into normal /grill-me mode for the rest of the session.