From 'Make It Cool' to 'Tighten the Letter Spacing': Requests AI Can Fix── Restate a vague adjective as a measurable attribute and a direction, such as narrower letter spacing, and the AI can change one thing you can then check

01One message after "make it cool," the operator named the letter spacing

In September 2026, the operator asked an AI to change the typeface on the English side of the site to something more modern and stylish. The same request added a side remark: the title on the entry page seemed a little too large.

In the next message, the request changed shape. The operator said the gaps between the letters looked too wide, and that a typeface with a tighter, more compact look would be better. A word of judgement, "cool," had been replaced by a property anyone can see: the space between letters.

The record keeps these two requests. It does not keep what the AI sent back to either one. Even so, the difference between the two requests is plain. The first does not say what would count as success. The second says where to make a change and in which direction.

This article makes one claim. A vague dislike, first felt as an adjective, becomes something an AI can fix once it is restated as a measurable property, which this article calls an attribute.

02An adjective judges the result; an attribute is the cause of that judgement

"Cool," "hard to read" and "heavy" are judgements made by whoever is looking. They differ from person to person and can shift for the same person from one day to the next. Hand one to an AI and it cannot feel the judgement for itself.

An attribute is the property that produces the judgement: letter spacing, font size, stroke weight, line height, colour depth. Each can be expressed as a number. Put two versions side by side and anyone can say which one has more of it.

There is no need to drop the adjective. The adjective says which direction you want to go. The attribute says which part to move, and by how much. A good request carries both.

03Without a measurable attribute, the AI cannot choose a target or a finish line

A request made only of adjectives costs you twice. First, the AI may change the wrong thing. There are many ways to make text look "cool": swap the typeface, thin the strokes, resize it, recolour it. If the AI picks a change that is not the one you cared about, the work has to be redone.

Second, there is no finish line. After the AI makes its change, you have to judge again whether the result is now cool enough. If that judgement wavers, the exchange runs to many rounds.

Anthropic's published prompting guidance says that being specific about the desired output improves results. It also suggests a test: show your prompt to a colleague who lacks context on the task, and if they would be confused, the model will be too. "Make it cool" is exactly the kind of phrase that would leave such a colleague guessing.

04Restating means fixing three things: attribute, direction and amount

In this series, restating does not mean swapping one adjective for another. "A tighter look" is still an adjective. But the operator's second message also contained an attribute and a direction: the gaps between letters were too wide. That is what turned it into a request the AI could act on.

You do not need the amount at the start. With an attribute and a direction, the AI can offer a small change. You look at it and say "more" or "go back." The amount settles on the second or third round.

Some adjectives do not reduce to one attribute. "Make it feel trustworthy" is a judgement built from several properties at once. In that case, do not try to fix it in one step. Ask the AI to list candidate attributes first, and choose which one to move yourself.

Once the request names an attribute, it can be checked against outside standards. WCAG 2.1, the international guideline for accessible web content, requires that content still work when a user widens letter spacing to 0.12 times the font size. A request to tighten spacing can say that it must stay within that condition.

05In material review, turning adjectives into attributes gets a fix done in one round

In pharmaceutical promotional material review, comments are often written as adjectives. The wording is too strong. The safety information does not stand out. The whole thing is hard to read. These are sound judgements from the reviewer. Pass them unchanged to an AI, though, and the AI has to guess where to make the fix.

A comment written as an attribute lets the person who revised the piece and the person who reviewed it check the result against the same measure. Did the superlatives drop from three to zero? Is the safety text now the same size as the efficacy text? The answers come out as yes or no.

The same holds when marketing or medical staff ask a generative AI to revise a draft. Before writing "make it clearer," add one line saying which attribute should move to make it clearer. The drafts the AI returns will then point in one direction.

06AI fills an adjective with an average reading, so it changes things you did not mean

A language model builds its answer by choosing the continuations that have most often followed similar words in the text it learned from. Told "make it cool," it lines up the changes most often linked with that word, in order of plausibility. Nothing in that word tells it what the screen in front of you is missing.

As a result, the AI tends to change several attributes at once. The typeface changes, and so do the weight and the size. If only the spacing bothered you, parts that were fine also move, and judging whether the page improved becomes harder.

Requirements engineering has known this problem for a long time. NASA's systems engineering handbook asks that requirements be written so they can be verified. It lists words such as "easy," "sufficient," "user-friendly" and "fast" as examples of terms that cannot be verified. A request to an AI carries the same burden: can anyone check whether it was met?

The operator's first message already showed the difference. The coolness of the typeface was an adjective. The title size in the same message had an attribute, size, and a direction, smaller. That part was a request the AI could carry out in one place from the start.

MDN's reference for letter spacing adds one more reason to name the amount. It warns that very large positive or negative values make words hard to read, and that a readable value depends on the font. Once spacing is the named attribute, that limit becomes something you can check.

07From tomorrow, add the visible fact and the attribute under every adjective

There are four steps.

1. Keep the adjective. "Cooler" or "easier to read" stays as the direction you want.
2. Add one visible fact. Write what caught your eye, such as "the letters look spread apart."
3. Name the attribute and direction. "Narrower letter spacing," "a smaller title." If you are unsure, ask the AI to list three candidate attributes and pick one.
4. Move one attribute per round. Put the result next to the previous version and decide whether to add more or roll back.

Step four keeps cause and effect clear. When only one attribute moved, you know what made the page better or worse. Move several at once and you lose that.