Verified against ChatGPT · 2026-07-30
Pick the right reasoning depth for the task before you type the prompt
A short decision pass for choosing between ChatGPT's fast/instant response mode and its extended-thinking mode for a specific task, plus how to phrase the request so it actually invokes the depth chosen instead of getting auto-routed to the wrong one.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Before sending the actual task, help decide which response mode this specific task actually needs, and how to phrase the request so it invokes that mode deliberately rather than however the model happens to auto-route it. THE ACTUAL TASK Building a two-week on-call rotation for a 6-person team that respects everyone's stated blackout dates, keeps consecutive on-call weeks at least 3 apart per person, and balances weekend coverage evenly. WHAT KIND OF WRONG ANSWER WOULD BE COSTLY HERE A rotation that silently double-books someone during their stated blackout week would only surface when they're already unavailable during their shift — expensive and embarrassing to fix after the fact. HOW MANY CONSTRAINTS OR MOVING PARTS Six people, five blackout-date constraints, a minimum-gap rule, and a fairness rule across two months — at least four rules interacting simultaneously. TIME SENSITIVITY Needed by end of day, but not in the next five minutes — there's room for a slower, more careful pass. SELECTION RULES Recommend extended-thinking, or a higher-tier reasoning mode if available and warranted, specifically when the task has multiple interacting constraints that have to be satisfied simultaneously — a scheduling problem with several hard constraints, a multi-step logical deduction, a plan that has to account for several dependent variables at once — since these are exactly the cases where a fast, single-pass answer is most likely to satisfy some constraints while silently violating others it didn't have the working-through space to check against each other. Recommend the fast/instant mode when the task is a lookup, a rephrasing, a short creative draft, or anything where a wrong first answer is cheap and quick to spot and correct in a follow-up — spending extended reasoning time on a task that doesn't actually benefit from it just adds latency without changing the quality of what comes back. Weight the recommendation toward extended thinking whenever the cost of being wrong is high and the mistake wouldn't be obvious on a quick read — a subtly wrong answer that looks plausible is far more dangerous than an obviously wrong one, since only the obviously wrong one gets caught and redone; if a wrong answer here would look fine and get acted on before anyone noticed, that alone is reason enough to slow down. Note that phrasing a request to explicitly ask for reasoning through a problem step by step, naming and checking each constraint individually, or working through multiple scenarios before answering tends to invoke deeper processing even within a single mode, so the actual prompt wording matters independently of which mode toggle is selected. If time sensitivity conflicts with the recommended mode, say so explicitly rather than picking one consideration and ignoring the other, since that's a real tradeoff to be decided, not something to resolve silently in the recommendation. OUTPUT FORMAT One paragraph: recommended mode, the specific reason from the rules above that drove the recommendation, and a suggested phrasing for the actual request that would invoke that depth of processing regardless of mode selection. If there's a genuine mode-versus-time tradeoff, name it explicitly as a decision to be made, not one made silently.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
GPT-5.1-era ChatGPT routes between fast and extended-reasoning processing based partly on automatic complexity signals in the request and partly on explicit mode selection, and the automatic routing is a heuristic, not a guarantee — a request that reads as simple on its surface phrasing but is actually a multi-constraint scheduling problem can get auto-routed to fast processing that produces a single coherent-looking pass without ever checking all the constraints against each other simultaneously, exactly the failure mode extended reasoning exists to reduce. The specific criterion of "multiple interacting constraints" rather than a vaguer notion of "hard problems" gives a concrete, checkable trigger: a fast single pass through a problem tends to satisfy constraints roughly in the order it considers them and can commit to an early choice that turns out to conflict with a constraint considered later, whereas extended reasoning's actual mechanistic advantage is exploring and backtracking across exactly that kind of interaction before committing to a final answer. Weighting the recommendation by whether a wrong answer would be obviously wrong or subtly wrong targets a real asymmetry in how errors get caught: an obviously wrong answer gets redone immediately at near-zero additional cost, while a subtly wrong one that looks fine can get acted on and only surface as a problem much later, when it's far more expensive to unwind. Naming that prompt phrasing itself — instructing explicit step-by-step constraint-checking — can invoke deeper working-through independently of the mode toggle matters because relying on the mode selector alone treats depth as a binary switch, when in practice the actual text of the request meaningfully shapes how much the model works through a problem before committing to an answer, regardless of which processing tier picked the request up.
Verified against
ChatGPT GPT-5.1 · 2026-07-30
Changelog
- 2026-07-30 — Initial publish, verified against ChatGPT GPT-5.1 mode routing.
Building this for real?
This is a free starting point. If you'd rather have what Scult builds built and running for your business, that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
