ChatGPT Operator: Prompt to Autonomy · 18 min · 130 XP
Conflicts, preferences, and what to leave out
Handle sources that disagree, state how you want answers, and cut the rest.
Real source material contradicts itself. An old note says the event starts at 2pm; a later correction says 3pm. The wrong behaviour is to pick one silently — and note that picking the newer one is still picking silently, even though it's usually right.
What you want is the conflict surfaced: "These two sources disagree on the start time. The 3pm note is dated later and describes itself as a correction. Which should I use?" Then you approve one, and the approved value is what gets used from then on. The rule generalises: when sources disagree, report and ask; don't resolve by rule of thumb.
Working preferences are the standing instructions that save you repeating yourself — short paragraphs, metric units, explicit unknowns, no summary at the end. Write three, then test them across two different requests on the same brief. Preferences drift, especially in long conversations, and the only way to know whether yours survived is to check rather than assume.
Then cut the brief down. Add some unrelated background, see how the answers change, and remove it again. The exercise teaches something durable: most of what feels like helpful context is noise that dilutes the parts that matter. Keep a fact only if you can say what it would change — that's the test, and it eliminates more than you'd expect.
Practice. Create an older note saying 2pm and a newer correction saying 3pm, and confirm the conflict is reported rather than silently resolved — then approve one and check the approved time is what gets used. Write three working preferences and test them across two requests, marking which were followed. Then pad your brief with unrelated background, compare the answers, and strip it back, giving a reason for every fact you kept.
Loading your workspace…