Prompt Engineering or selling your soul?
If you use AI at work, you’ve probably had this experience:
You write a prompt that feels clear enough, hit enter, and the model confidently produces something that is… technically fine… but not what you meant.
Maybe it built the right thing in the wrong style.
Maybe it chose an approach you would never ship.
Maybe it made assumptions you didn’t realize you were leaving unstated.
I’ve started treating prompts like a contract with the devil.
Not because I think AI is evil — just because it’s literal, opportunistic, and perfectly willing to sprint in the wrong direction if you give it even a small opening.
And you can’t cover every edge case up front.
So here’s the one sentence I add to a lot of my prompts that consistently makes the results better:
Before you begin, ask any clarifying questions you need to fully understand what I’m asking and to do an excellent job.
Why it works (and why it’s “vibe engineering”)
Most prompting advice is basically: be more specific.
That’s true, but it’s incomplete — because the whole problem is that you often don’t realize what you forgot to specify.
This sentence flips the dynamic:
- Instead of “I describe something and hope the AI guesses right,”
- it becomes collaborative.
It turns the model into a reviewer before it becomes an implementer.
And it forces the “unknown unknowns” to show up early, while it’s still cheap to correct.
The kind of questions that save you
My favorite clarifying questions are the ones that expose missing context I didn’t realize mattered.
Like:
- “Is there an existing system you want me to use as a template?”
- “Is this a large table / high-traffic database?”
- “Is this safe to run during business hours?”
- “What does success look like: correctness, speed, low risk, or minimal code changes?”
- “Do you care about test coverage, or just a working fix?”
Those questions aren’t just helpful to the AI.
They’re helpful to me.
Because half the time, I’m using the AI to tease out details I forgot to include in the prompt in the first place.
A realistic failure mode this prevents
Say you ask:
“Write a migration to backfill X safely.”
The AI might happily generate a perfectly valid migration that:
- locks the table,
- runs as one big transaction,
- does a full scan,
- adds an index in a way that takes forever,
- and generally assumes the world is a small quiet sandbox.
It’s not “wrong” — it’s just not something you want to discover after you already committed to the approach.
If you make the model interview you first, you’ll often get the question you forgot to say out loud:
“How large is the table, and can this run on prod without blocking writes?”
That one question can save you a very annoying afternoon.
When I don’t use it
I don’t paste this sentence into every single prompt.
If I’m asking something small and obvious (“write this one-liner”, “explain this error”, “rename these variables”), it’s unnecessary overhead.
But if I’m doing any of these, it comes out almost automatically:
- writing something new (a real feature, not a snippet)
- debugging a complex problem
- anything with multiple moving parts
- anything where “technically correct” can still waste time
Copy/paste template you can steal
Here’s a version you can drop into your own prompts:
My default prompt preamble
Before you begin, ask any clarifying questions you need to fully understand what I’m asking and to do an excellent job.
Optional follow-up (if you want to be extra explicit)
- If something is ambiguous, don’t guess — ask.
- If there are multiple valid approaches, list the options and tell me what you recommend and why.
- After questions are answered, produce the output in a clean, usable format.
That’s it.
It’s not magic. It doesn’t make the model smarter.
It just keeps the process collaborative instead of “write a prompt and hope for the best.”
Teaser: if you want to hear about the #2 sentence I add to prompts to make them even better, get me to 100 likes 😄
Top comments (0)