4 prompts that unstick a stalled ChatGPT session (and 1 that doesn't)
You've been talking to the model for forty minutes. The last few answers were OK, not great. You can feel it hedging. You also can't remember what to say next, because the thing you actually want is for the model to stop doing the thing it's been doing.
This is the moment Keep Working was built for.
We ran a blind evaluation: 345 carefully written nudges versus two baselines: are you sure? (three words, no curation) and Review your last answer for errors and improve it. (the well-meaning version of the same thing). Three independent LLM judges, randomized order, no idea which prompt produced which output.
Four categories of prompt beat both baselines. Here they are.
1. Step back, before going forward
Wait. Are we even solving the right problem?
This one cleaned house. On the most ambiguous snippet in our eval, the redirect-stepback category went 3/3 in blind comparisons. The trick is that "are you sure?" makes the model recheck the work; redirect-stepback makes it recheck the problem.
Use it when the answer seems competent but you can feel the goal has drifted. The model will frequently come back with: "Actually, I think you wanted X instead." That's worth more than three more paragraphs of the wrong answer.
2. Tear your own answer apart
Now tear that apart. Where is it weakest?
The adversarial-critique category went 3/3 on the snippet it was tested against. This is the most direct prompt in the deck: it cancels the model's incentive to look good and replaces it with the opposite incentive. The output reads like a critic, not like an apologist.
Use it after any substantive answer. The interesting weakness is rarely the one the model would surface unprompted.
3. Force a single decision
Name the actual recommendation, the one criterion that decided it, and the strongest counter-argument you rejected.
The finalize-ship category went 3/3 with confidence 4 from every judge. This is the prompt for when the model has been "menu-ing": listing four options without picking one. The structure forces it to commit, then justify, then concede.
This is the only category our eval said to ship as a standalone card before the rest of the deck stabilized. It worked that well.
4. Pin the numbers
Pin the numbers. If you said "large", "slow", "many", or "most", give the actual figure — or the order of magnitude you actually mean.
The concretize category was bimodal: it won when the underlying weakness was vagueness, lost when the underlying weakness was structural. On the snippets where vagueness was the actual issue, concretize hit hard.
Use it when the answer reads smooth but you couldn't act on it tomorrow. Specifically: numbers, names, units, citations. The model usually has more specificity available than it volunteers.
The one that didn't work
stress-test, the category that asks the model to walk through edge cases, went 0/3 in the v1 eval and got worse in subsequent iterations. The category survives in the deck only as an experimental opt-in because the generic baseline "what could go wrong?" out-performed every specialized variant we tried. Sometimes the prompt that works is the prompt that doesn't try too hard.
Why these four
The pattern is that they're forcing functions, not scenarios. Each prompt creates a single, clear demand (pick one, restate the problem, tear it apart, pin the number) that the model can't dodge. The prompts that lost in our eval tended to describe a scenario ("two users hit this at the same millisecond"), which gave the model an excuse to answer that scenario instead of the actual flaw.
If you want all 86 of the core prompts, the deck is at keepworking.urbanodx.com. Space for next, C to copy, F to flip.
For the long story of how we got here, including the v1 deck losing decisively to three words, see how 'are you sure?' beat 345 hand-curated AI prompts.