From Asking Questions to Engineering Prompts
Context
After using ChatGPT to support real interview deliverables, I ran into an uncomfortable pattern.
Sometimes the output was sharp, structured, and genuinely useful.
Other times it was fluent, confident, and wrong.
Nothing obvious had changed. Same tool. Similar questions. Same pressure.
That inconsistency forced me to stop blaming the model and look at my own inputs.
Problem being solved
I believed that better results came from asking better questions.
That belief did not survive contact with reality.
What I was actually doing was delegating thinking without instructions.
I was leaving gaps and letting the AI decide how to fill them.
It did exactly what it was designed to do:
produce plausible output, even when the framing was weak.
The problem was not that ChatGPT misunderstood me.
The problem was that I had not decided what “good” looked like before asking.
How ChatGPT was used
At this stage, ChatGPT stopped being treated like a search engine or an explainer.
I stopped expecting answers.
Instead, I started treating it as a system that reflects the quality of its inputs.
If the framing was vague, the output drifted.
If the constraints were missing, the assumptions multiplied.
Its role shifted from “tell me” to “work within this frame”.
That single change altered everything.
What was built
There was no formal output at this stage.
No documents. No reusable assets. No deliverables worth sharing.
What was built was internal.
I started shaping inputs deliberately:
deciding the role before the request
clarifying context instead of implying it
defining what the AI was not allowed to do
specifying the shape of the output, not just the topic
The value was not in the text ChatGPT produced.
It was in the thinking that happened before I typed the first word.
What failed or broke
Several assumptions broke quickly.
Vague prompts produced confident nonsense.
Overloaded prompts diluted focus.
Missing audience context created unusable tone.
Trusting the first response wasted time later.
The most dangerous failure was subtle:
good sounding output hid bad thinking.
The AI did not challenge weak framing.
It amplified it.
What insight emerged
Prompting is not about asking questions.
It is about engineering conditions for useful work.
The model did not become smarter.
The inputs became clearer.
Structure did not reduce creativity.
It constrained chaos.
Once I accepted that output quality was my responsibility, consistency stopped being a mystery.
Why it led to the next phase
Once outputs became more predictable, a new constraint appeared.
Scale.
If this way of working mattered, it could not live only in my head.
It needed to be explainable.
Repeatable.
Teachable.
That requirement forced the next step: formalizing what had been implicit.
Internal links
Back → Post 1: How I Used ChatGPT to Build Real Job Interview Deliverables
Forward → Post 3: Building and Stress-Testing a Prompt Engineering Training Manual
