A vague prompt is not a shortcut. It's a loan at 400% interest.
Ten seconds saved typing. An hour lost fixing. And the worst case isn't the hour: it's the confident, plausible, wrong output that nobody catches until an operator does.
THE BRUTAL TRUTHEvery word you don't write, the model guesses. Every guess you don't check, ships. "Make it better" has cost this company more hours than any outage, and nobody logs it because it feels like the AI's fault. It isn't.
The cost ladder
A lousy prompt doesn't cost one thing. It costs four, in order, and each rung is worse than the last. Most people only ever notice rung one.
1
Your time. The model guesses the audience, the format, the scope. You read 600 words of wrong, re‑prompt, read again. Three to six rounds.typical: 40–90 min for a task that needed 8
2
The reviewer's time. Whatever you didn't specify, someone else now has to spot. Ondrej, the tech lead, compliance. They read plausible text and hunt for what's silently wrong. That is the most expensive kind of reading there is.typical: 20–40 min of a senior person, per document
3
The thing that ships wrong. Release notes to an operator with the wrong RTP variant. A Jira fix that "worked" and broke the API layer. A LinkedIn post promising something Legal never approved. Now it is an incident, not a prompt.typical: hours to days, plus an apology email
4
Trust. After the third incident, the team stops trusting AI output. Not your output: all AI output. The company loses the leverage. That is the real cost, and it is on you.typical: months
10 s
to type "fix the lobby bug"
2 min
to type the six‑line version
~1 h
difference in total time, in the wrong direction
Real THG examples: lousy vs. specific
Each pair is a real shape of a real request that went through this company. Left is what got typed. Right is what should have been typed. Load either one into the grader below and watch the score.
The lousy prompts are not wrong. They are underspecified. The model will still answer, fluently, confidently, and it will fill every gap with the statistically average guess. The average guess is never THG‑specific. That's the trap.
The sandbox: type a prompt, watch it get graded live
This grader is deliberately blunt. It checks for the things a model cannot guess: who it's for, what exactly, what it must not do, what shape the answer takes, and how you'll know it's right. It also hunts vague words. Type, and the grade updates as you go.
Start from a lousy preset. Push it to an A without padding it with fluff. Then press Run it and watch what the afternoon looks like at each grade.
Prompt grader
0 words
what this prompt costs · simulated
grade your prompt, then press ▶ Run it
F
…
score 0/100
Quiz: prove it
Five questions. Four right to pass. The progress record only stores passes.
Module 2 quiz
THE ONE THING TO REMEMBEREvery word you don't write, the model guesses. Every guess you don't check, ships. Two minutes of specifics is the cheapest time you will spend all week. Spend it.