TL;DR
Left uncontrolled, every AI answers in the same shape: the medium-length essay — safe, padded, and mostly filler. Scope limits tell the model what not to cover, and format constraints dictate exactly what shape the answer must take; together they convert rambling text into dense, usable output.
The Analogy
Imagine hiring a master chef, and the only instruction you give is: "food."
The chef is delighted — and dangerous with a budget. Out comes a twelve-course tasting menu: soups you didn't want, a palate cleanser, three desserts, a monologue about reductions. Every course is technically excellent. The bill is enormous. And somewhere in course nine, you realize you just wanted a sandwich.
Now watch the same chef receive: "One grilled cheese sandwich, sourdough, cut in half, on a small plate." No tasting menu. No monologue. Thirty seconds of pure, sandwich-shaped competence.
An unconstrained prompt is "food." Limiting scope is choosing the dish instead of the menu. Forcing the format is specifying the plate. The AI, like the chef, is built to impress — and with no boundaries, "impress" means "more." Words, sections, caveats, options, an essay-shaped parade of effort you will pay for line by line and read almost none of.
Here's the detail most people miss: constraints don't restrict quality, they concentrate it. Ask for twelve courses and the average quality per course drops; ask for one perfect sandwich and every ounce of skill lands between two slices of bread. The chef was always this good. You finally aimed him.
How It Works & The Economy (Uzu)
Two dials, both turned in the prompt itself.
Dial one — scope: define what's out. A model facing an open question hedges in every direction at once, because covering everything is the safest average answer. Close the exits: name the audience, the purpose, the depth, and what to exclude. This is half of the one-shot prompt anatomy — the context component — doing its quiet work.
Dial two — format: define the shape. A format constraint ("as a table", "exactly 3 bullets", "max 100 words", "one paragraph, no preamble") does something subtle: it gives the model a fixed container to fill, so its effort goes into substance instead of structure. You can feel the difference in the first sentence — constrained answers open with the answer, not with "Certainly! Here's a comprehensive overview..."
| Constraint type | What you write | What it kills |
|---|---|---|
| Length | "max 100 words" / "exactly 3 bullets" | Padding, preambles, recap paragraphs |
| Structure | "as a table" / "numbered list" / "one paragraph" | Essay drift, buried ledes |
| Exclusion | "no introduction, no conclusion" | Ceremony and throat-clearing |
| Depth | "for a beginner" / "expert level" | Both over-explaining and under-explaining |
And the constraints compound where it counts: a tight answer is a small answer, and small answers leave more room on the context window — the desk where every follow-up re-reads everything. A chat built from three-bullet answers stays sharp for dozens of turns; a chat built from essays drowns its own instructions by message fifteen.
This is also where Uzu makes the physics visible: credits. Generating a long, drifting answer costs far more than generating a tight one, so forcing a compact format isn't just readability — it's a financial strategy. Every word the model doesn't write is a fraction of a G-Credit you keep. The prompt that says "exactly 3 bullets, max 20 words each" and the prompt that says "tell me everything" may both be 'correct', but they burn wildly different amounts of your economy — and only one of them produces an answer you'll actually use.
So the working rule is simple:
Constrain first, then ask. If you can describe the shape of the ideal answer, put that description in the prompt — the model fills containers far better than it guesses them.
Before & After (The Constraints)
Example 1 — the open question
Explain photosynthesis.
Four paragraphs, a section on history, two metaphors, and a summary. Beautiful, and triple the length you needed.
Explain photosynthesis to a 10-year-old in exactly 3 sentences. No analogies involving factories.
Same knowledge, compressed into a container a child can hold. Note the exclusion — "no factory analogies" — closing an exit most answers would happily wander through.
Example 2 — the comparison
Compare solar and wind energy.
A balanced essay with six subtopics, hedged conclusions, and the key numbers scattered across paragraphs nine and eleven.
Compare solar vs. wind for a home installation as a table with exactly 3 rows: upfront cost, payback time, space needed. One-word verdict after the table.
The table forces the comparison into parallel facts; the verdict constraint forces a decision. Nothing survives that doesn't earn its row.
Example 3 — the rewrite
Improve this email.
The model returns a longer email — because "more thorough" is its default definition of "better."
Rewrite this email under 80 words, keeping the first line as the ask, bulleting the two logistics, no sign-off change.
Length capped, structure dictated, scope locked to this email. The result is shorter than the original and better, which no default setting would ever produce.
Common Pitfalls
- Assuming the AI knows when to stop. It doesn't — models have no sense of "enough," only a sense of "complete." Without a limit, "complete" grows toward essay. Always attach a size to the deliverable.
- Constraining the format but not the scope. "As a table" applied to an enormous question yields an enormous table. The two dials work together: shrink the territory, then choose the container.
- Setting limits you don't actually want. "Max 50 words" when you needed nuance produces a clipped, distorted answer. Constraints should mirror the real shape of your ideal deliverable — if you'd accept 200 words of genuine insight, don't demand 50.
- Forgetting negative constraints exist. "No preamble," "no emojis," "don't mention pricing," "skip the history" — beginners only describe what they want, never what to leave out. Exclusions are the sharpest scope tool in the box.
- Saving the constraints for the follow-up. Sending "shorter, please" after a bloated answer means the tokens were already generated — and in Uzu, already paid for. Constraints are cheap before generation and expensive after it.
FAQ
How do I make AI give shorter answers?
Attach an explicit, measurable constraint to the prompt itself: "in exactly 3 bullets," "under 100 words," "one paragraph, no preamble." Vague requests to "be brief" fail because they have no test — "exactly 3 bullets" can't be misinterpreted. Add negative constraints ("no introduction, no conclusion") to trim the ceremony models add by default.
What does limiting scope mean in prompting?
Limiting scope means defining the boundaries of the answer before it's generated: the audience, the depth, what's included, and — most powerfully — what's excluded. It's the difference between "tell me about marketing" and "3 Instagram caption ideas for a bakery, under 20 words each." Same model, wildly different density.
Does a shorter AI answer save money?
Yes. Language models generate text token by token, and generation is the expensive part — a 900-word answer costs roughly nine times a 100-word one. In Uzu, that difference is charged in G-Credits, so a forced-tight format directly stretches your balance: constrain the shape first, pay for substance instead of padding.
Next Lesson
So far every technique has been about words in, words out. But modern models have another sense entirely: they can see. Feeding them an image changes the rules of prompting — vocabulary replaces syntax, and the cost model changes too. Welcome to the most misunderstood input in AI.
Continue to the next lesson: Visual Prompts & Attachment Costs
Practice drill before you go: take the last long AI answer you received and reverse-engineer the constraint that would have produced it at one-third the size — the exact word count, the exact container, the one exclusion that would have cut the filler. Next prompt, write that constraint first.