Vague versus precise
The same task, written two ways. The vague version may look plausible, but it can still be wrong in very specific ways because important decisions were never stated.
"Make me a nice pricing page" produces a pricing page. It will look fine. It will also invent your prices, pick three tiers because pricing pages have three tiers, claim one is popular, and use a grey that fails contrast.
None of that necessarily means the model failed. Those decisions still had to be made, and because you did not specify them, the model filled them in for you.
A specification is not a longer prompt. It is the list of decisions you refused to delegate.
Work the instrument below. Each invention on the right is something the model chose; click it and watch it move into your prompt. Do all six and the two sides converge.
A spec is the decisions you did not delegate
0 of 6 taken backMake me a nice pricing page.
All the features you would expect
Get StartedAll the features you would expect
Get StartedAll the features you would expect
Get StartedBoth results are hand-written, not generated. No model runs on this page; the shape of the difference is the lesson, and you run the real thing yourself in a moment.
The page on the left is not broken. It is plausible, and it is specifically wrong in six places, and every one of them is a decision nobody made on purpose. Click one to take it back.
Count the result: six lines. Those six lines can be the difference between a page you would ship and one you would not. The improvement comes from making the decisions explicit, not from writing cleverer English.
From here on, the move before any significant prompt is thirty seconds long: what has to be true about this that I have not said?
You should now be able to
- Tell a wish from a specification
- Find the decisions you handed over without meaning to
- Write a prompt that produces something you can use
Loading…