The "interview me first" prompt
Stop writing three paragraphs of context nobody asked for. This prompt makes the model interview you before it starts, and the first draft comes back needing one revision instead of four.
- Use case
- Writing & planning
- Works with
- Claude, GPT-class, local 14B+ (cap it at 5 questions)
I want your help with a task, but do not start it yet.
Task: {describe the task in one or two sentences}
First, ask me the questions you need answered to do this well. Ask up to
seven, in order of importance, and after each one tell me in a few words
why the answer matters. If one of my answers changes what you would ask
next, ask the follow-up.
Only start the actual work after I have answered. If I skip a question,
make the safest assumption and mark it ASSUMPTION in your output so I can
spot it.
Every bad first draft I get out of a model fails the same way: it filled the gaps in my request with the most average possible guesses. Wrong audience, wrong length, wrong level of formality. Not because the model is dumb, but because I never said.
This prompt flips the setup. Instead of me trying to predict what context the model needs, it tells me. The questions it asks are consistently better than the context I would have volunteered.
Why it works
- “Do not start it yet” is doing the heavy lifting. Without it, models answer the task and tack a question on the end as decoration.
- “Why the answer matters” filters the questions. It stops the model from padding the list with things it does not actually need.
- The ASSUMPTION marker means a skipped question costs you nothing. You scan the output for the word and either accept the guess or correct it.
What good output looks like
You describe “an email to my ISP about a billing dispute” and it asks how much money is at stake, whether you have prior case numbers, and whether you want resolution or escalation. Three questions you would not have thought to answer, all of which change the email.
Where it works (and doesn’t)
Claude and the GPT-class models handle the follow-up branching well. Local models around 14B follow the structure but ask flatter questions; cap them at five or they wander. Below 8B, skip this one, the model tends to answer its own questions and barrel ahead anyway.
The five minutes of answering questions replaces the forty minutes of revision rounds. That trade wins every time.