Before agents, before MCP, before every model maker started racing to out-benchmark the last, there’s only one word that could win the Guinness World Record for the longest-serving term in the AI dictionary: prompt. Everything else has come and gone or been rebranded three times over, but the prompt has stayed exactly where it started. It’s still the single thing standing between you and a good answer. Sure, the prompt engineering workshops and the $200 master ChatGPT courses have mostly faded out, and half the people who called themselves prompt engineers in 2023 have quietly changed their LinkedIn titles since.
However, the thing they were all circling never went anywhere, and the word stuck around because the job its name never actually got automated anyway. No matter how capable the model gets, it still can’t read your mind about what you actually wanted. You could be on the highest tier plan, burning through API credits like you’re Sam Altman, running the most capable model Anthropic has ever shipped, and still get an answer you’d swear came from a model from the early AI days. So a while ago, I started adding a single line to the end of my prompts, and the improvement was immediate.
Claude keeps answering questions I don’t actually ask
The confidently-adjacent response problem
No one’s surprised by AI hallucinations at this point, and everyone knows well enough to take anything an AI says with a healthy pinch of salt. While AI labs have been directing their efforts towards ensuring their newer models hallucinate less and less, there’s a failure mode that no amount of benchmark-chasing seems to fix: answering the wrong question correctly. Now, Claude isn’t making anything up here and its giving perfectly accurate responses. AI tools have gotten significantly better at that. However, the answer is just accurate to a question slightly different from the one I meant to ask.
Since nothing about it is technically wrong, it never trips the alarm a hallucination would. As a consequence, LLMs answering questions you never asked has never really gone away, and it’s still something you’ll run into fairly often no matter which model you’re on. Now, the reason why this happens comes back to the prompt you use. Think about it: if you ask someone a question where you don’t give them the full picture, they’ll fill in the gaps themselves with their own assumptions, their own context, their own best guess at what you meant. A human at least has the option to stop and ask you to clarify. And while I find Claude and other AI tools often stop to get some clarification, it typically happens only when what you’re asking is extremely vague.
For example, if I ask Claude to build a webpage for me, it’ll almost always stop and ask me what type of layout I want and what my aesthetic preferences are. Ask for something that sounds specific enough, like write me an essay on the French Revolution, and Claude won’t pause at all since the request reads complete. It’ll just begin, quietly making a handful of decisions on your behalf as it goes. It’ll pick a length, even though “an essay” could mean three hundred words or three thousand. It’ll settle on a reading level and tone — is this for a high schooler, a university seminar, or just you trying to understand the thing over coffee? Claude will sometimes just decide itself. The same goes for when you ease your way into a topic without ever spelling out what you actually want. It sometimes doesn’t wait to find out, and just picks a lane and drives.
This one line fixed most of it
Twelve words doing what my longer prompts couldn’t
The fix that has worked for me most consistently isn’t a clever prompt template or some 300-word system message I copied off Reddit. It’s a single sentence I paste at the end of whatever I’m asking (and edit a bit as needed): Before you answer, tell me what you need to know to answer well, and point out any assumptions you’d otherwise make.
With this line at the end of every prompt, it nudges Claude to lay out what it doesn’t know yet. Going back to the French Revolution essay, Claude won’t produce a single word of the essay with the line appended. Instead, it asks how long I want it, who it’s for, whether I want a narrow argument or a broad overview, how formal it should be, whether I need sources, and so on. Essentially, this prompt hands me the exact list of decisions it would have made silently, and lets me make them myself. The same five guesses that used to happen invisibly now happen out loud, before anything gets written.
Now, Claude might’ve stopped and asked me a few questions if I hadn’t used this prompt. However, given I was upfront with what I wanted and clearly mentioned I wanted assumptions ironed out before anything got written, it asks far more than it otherwise would’ve. Left to its own judgment, Claude only surfaces the one or two things it considers obviously unclear.
Claude Projects is the best productivity booster that you’re not using
It checks a lot of boxes for me.
With the line, it surfaces the whole set including the quiet decisions it would normally feel confident enough to just make on its own. Those are exactly the ones that tend to send an answer off in the wrong direction. Without the prompt, I’d have needed to catch every one of those decisions myself. This would have been in the form of either cramming every detail into the original request up front, which assumes I even know what matters before I see the answer, or noticing after the fact that Claude quietly went the wrong way and chasing it with constant follow-up messages. Both technically work, but cost me the exact same thing I came to Claude to save: time.
This works with any AI tool you already us
Of course, you can use this prompt with any LLM you use, whether it’s a cloud-based model or a local LLM. There’s no rocket science to it, and the line is just meant to nudge the model into showing its working before it commits to an answer.
That habit, filling in gaps you never specified and sounding confident while doing it, isn’t a Claude quirk; every model does it. So whether you’re on ChatGPT, Gemini, a local Llama model, or Claude, the same sentence pulls the same silent guesses out into the open before they have a chance to steer the answer wrong.
