How to Add Constraints — Prompting 101

You'll learn

What constraints add that the four parts don't
The constraint families: length, tone, format rules, and things to avoid
How to describe a wrong answer so the AI can steer away from it
How to constrain without strangling the output
A constraint menu you can append to any prompt

01Section

Boundaries, not instructions

The four parts — role, task, context, format — point the AI at the target. Constraints do the opposite job: they fence off the misses. An earlier lesson called them "telling the AI what a wrong answer looks like." This lesson turns that one line into a full skill.

Here's why they matter: every AI tool ships with defaults. Left alone, you get medium-length answers, a relentlessly cheerful tone, an essay shape, and a fondness for phrases like "in today's fast-paced world." A constraint is how you override a default you don't want.

Plain-language definition

Constraint — a boundary you set on the answer — how long, what tone, what to include, what to leave out. Constraints don't tell the AI what to do; they narrow what it's allowed to hand you.

A useful analogy

Ordering "a coffee" gets you the house default. Ordering "a small oat-milk latte, no sugar" gets your coffee. The barista was capable of both all along — the constraints are the difference.

Key takeaway

Generic output isn't only a context problem. Sometimes the AI understood you fine — and gave you its defaults anyway. Constraints override the defaults.

Checkpoint

02Section

Length and quantity: use numbers

The most common complaint about AI output is that there's too much of it. The fix is a number. "Keep it brief" means one thing to you and another to the AI — "under 100 words" means the same thing to both of you.

The same goes for quantity. Ask for ideas and you'll get a dozen padded ones; ask for "exactly three, one sentence each" and every idea has to earn its slot.

Adjective — the AI decides what it means

Write a short summary of this report. Keep it brief.

Numbers — no room to interpret

Summarize this report in exactly five bullet points, each under fifteen words.

The rule

Numbers beat adjectives. "Under 100 words" beats "brief." "Exactly three options" beats "a few ideas." "Two paragraphs" beats "not too long."

Checkpoint

03Section

Tone: give a reference point

Tone words work at the extremes — "formal" and "casual" land reliably. The middle ground is where they blur: "friendly professional" produces something different in every tool, on every day.

Two moves are far more reliable. First, a reference point: name something whose tone the AI already understands. Second, a banned list: name the tics you don't want, because you'll otherwise get them.

Tone by adjective

Write it in a friendly but professional tone.

Tone by reference point + banned list

Write it like a short note to a colleague you like — warm, direct, plain. No exclamation marks, no "I hope this email finds you well," no "excited to announce."

The reference point ("note to a colleague you like") does what three adjectives can't, and the banned list blocks the tics that would sneak in anyway.

Quick knowledge check

You want a specific tone and "friendly professional" keeps missing. What's the most reliable fix?

ARepeat the adjective in capital lettersBGive a reference point and list the phrases to avoidCAsk for the tone in a separate prompt afterwardsDSwitch AI tools until one guesses right

Checkpoint

04Section

Say what a wrong answer looks like

The most underused constraint is the negative one: naming the failure you've seen before. If your last three AI drafts all opened with a rhetorical question, or padded every list with obvious filler, don't hope the next one won't — say so.

This is also where accuracy rules live. An AI tool will happily fill a gap in its knowledge with something plausible, so close that door explicitly.

An accuracy constraint worth reusing

Do not invent statistics, studies, quotes, or sources. If you don't have the information, say so instead of guessing.

Don't over-constrain. Fifteen rules make the AI optimize for rule-following instead of doing the job well — pick the three or four that matter.
Watch for contradictions. "Be thorough" plus "under 50 words" forces the AI to quietly drop one of them, and you won't know which.
No constraint guarantees the output. "Don't invent facts" reduces fabrication; it doesn't eliminate it. Anything important still gets checked by you.

Quick knowledge check

Your last three AI drafts all opened with "In today's fast-paced world..." What's the best fix?

ADelete the opening by hand every timeBAdd a constraint naming it: "Don't open with a scene-setting cliché — start with the point."CWrite longer prompts with more contextDAccept it as how AI writes

Checkpoint

05Section

A menu you can append to any prompt

You don't need to invent constraints from scratch each time. Keep a short menu and append the two to four lines that fit the job — the brackets show where your specifics go:

The constraint menu

Keep it under [NUMBER] words.
Give exactly [NUMBER] options, [one sentence] each.
Format it as [a bullet list / a table / short paragraphs].
Write it like [a note to a colleague / a plain-language summary for a customer].
Use plain language — no jargon, no buzzwords.
No exclamation marks, no "I hope this finds you well," no "excited to announce."
Don't open with a scene-setting cliché — start with the point.
Do not include [thing you keep getting that you don't want].
Do not invent statistics, studies, quotes, or sources — if you don't know, say so.
If anything is unclear about what I'm asking, ask me before answering.

Treat it as a menu, not a checklist — appending all ten lines to every prompt is exactly the over-constraining the last section warned about. Pick the lines that block the failures you actually expect.

Pause and think: Which AI default annoys you most — the length, the tone, or the shape? That's the first constraint to start appending.

Checkpoint

Prompt exercise

Same task, with and without boundaries

This prompt asks the AI to do one task twice — once with its defaults, once inside your boundaries — so you can see exactly what constraints buy you. Copy it into ChatGPT, Claude, Gemini, Copilot, or whichever AI tool you have access to; the website doesn't run AI itself.

I want to compare unconstrained and constrained output. First, ask me three quick questions, one at a time, to gather details for a short "about me" paragraph for a personal website. Then write it twice. Version 1: no special rules — write it however you naturally would. Version 2: 60 to 80 words, first person, warm but plain, no buzzwords like "passionate" or "driven" or "innovative," no exclamation marks. Label both versions clearly.

Reflection: Put the two versions side by side. Which constraint earned its place — and was there a default the AI chose in version 1 that was actually fine?

Quick check

3 quick questions — no pressure

There's no pass or fail here. Answer them all, and we'll show you the answers either way.