Say what good looks like

“Make it more professional” is one of those phrases that feels like an instruction and behaves like a mood. Ask three colleagues to apply it to the same paragraph and you’ll get three different paragraphs, all defensible.

The model has the same problem, with less recourse — it can’t ask you what you meant, so it picks the most common reading of “professional” and hands it back. If that isn’t yours, you go around again.

Adjectives are ambiguous; specimens are not

The fix is unglamorous: show one line of what you want.

Rewrite these six product descriptions to match this style: “Kettle, 1.7L. Boils in three minutes. Stainless steel. £34.” Short, factual, price last, no adjectives about how lovely it is.

That specimen settles length, tone, punctuation, ordering and vocabulary in a single stroke — and it settles them the way you meant, not the way the word “professional” is usually meant. You will spend fewer words writing the specimen than you would spend arguing toward it.

This works because of what L1 established. You’re not describing a target; you are beginning a document whose pattern the model continues.

Make constraints checkable

A constraint you can’t verify is a wish. Compare:

Wish	Constraint
”Keep it concise"	"Under 200 words"
"Make it scannable"	"Six bullets, one line each"
"Sound friendly"	"Second person, no exclamation marks"
"Be thorough"	"Cover all five sections; say so if a section has no data”

The right-hand column has a property the left-hand column lacks: you can look at the output and know whether it complied. That turns a vague dissatisfaction into a specific correction, which is what makes the next round fast.

Shape as an error detector

There’s a second reason to specify format, and it’s the one people miss.

Ask for prose and a fabricated detail hides comfortably inside a fluent paragraph. Ask for a table with a column headed “where this came from” and a fabrication has nowhere to sit — the cell is either filled with something you can check, or conspicuously empty.

You are not making the model more truthful. You are making its output easier to audit, which is the part you control.

The limit worth knowing

Format compliance is good, not perfect. Long outputs drift: a model asked for exactly ten items sometimes returns nine or eleven, and word counts are approximate because it is predicting text, not counting it.

Treat counts as targets rather than guarantees, and check anything that matters.

Try it now (6 minutes)

Take something you rewrite often — a status update, a meeting note. Write one specimen of the version you actually want. Then ask for five more in that shape.

Save the specimen. That’s the seed of the reusable prompt library you’ll build in L8.

Check your understanding

1. 'Make it more professional' underperforms because: The model ignores adjectives It's ambiguous, so the model applies the most common reading rather than yours Tone can only be set in paid tiers

2. Which is a checkable constraint rather than a wish? Keep it punchy Six bullets, one line each Make it feel modern

3. Asking for a 'where this came from' column mainly helps because: It makes the model incapable of inventing It gives fabrications nowhere to hide, making output easier to audit It forces a web search

Recap

Show a specimen instead of stacking adjectives, prefer constraints you can check, and choose shapes that make errors visible. Counts are targets, not guarantees. Next: the follow-up, where most of the real quality comes from.