ChatGPT Operator: Prompt to Autonomy · 18 min · 130 XP

Source-bound answers, decision logs, and handoff

Bind answers to your material, record what was settled, and start fresh safely.

When you supply a document and ask a question it doesn't cover, there are two possible answers: what the document says, and what venues generally do. The second is often useful and is never what you asked for. Bind the answer to the source: if the policy doesn't state it, the answer is "not stated" — optionally followed by general advice, clearly marked as such.

Test this deliberately. Supply a fictional venue policy, ask something it can't answer, and see whether you get the honest gap or plausible general guidance in the voice of your document. It's a two-minute experiment that tells you how much to trust every other answer about that document.

A decision log is the brief's other half. The brief says where you're going; the log says what's been settled on the way. Four columns: the decision, the reason, the date, and what's still open. Then audit it — the failure mode is a suggestion from a conversation appearing as an approved decision. You wrote "we could do evenings"; the log says "evenings confirmed".

Handoff is the payoff. Open a brand-new conversation with nothing but the brief and the log, and ask for the next three actions. If they're consistent with your recorded decisions and assume nothing you didn't write down, the pair is doing its job. If the answer needs something you never wrote, you just found the gap — and you found it in a new chat rather than in a month when you'd forgotten the context entirely.

Practice. Supply a fictional venue policy and ask a question it cannot answer, requiring an explicit "not stated" instead of general advice. Turn a practice conversation into a decision log with decision, reason, date and open question, then check nothing was promoted from suggestion to decision. Finally, start a fresh chat with only the brief and the log, ask for the next three actions, and confirm none of them assumes context you never supplied.

Loading your workspace…