Prompting 101: clear asks, roles, constraints

You open a new chat. You type, “Summarize this.”
The model dutifully spins paragraphs you didn’t ask for, forgets the bullets you wanted, and somehow adds a motivational quote. It’s not trying to be difficult—it’s doing what language models do:

This guide teaches you how to give instructions the model can’t misunderstand. You’ll learn a compact mental model, see it applied end-to-end, and practice with a short lab. By the end, you’ll hit target formats—bullets, tables, steps—without wrestling the output.

LLMs are excellent at following structure and mediocre at guessing your intent. If you say “summarize,” the model invents a structure. If you say “five bullets, last bullet a risk,” it stops guessing and starts following.

Two quick definitions so we’re on the same page:

• Prompt: Everything you send the model in the conversation, including system and user messages.

• Context window: The model’s “short-term memory.” Put critical instructions up front so they’re hard to miss.

💡 Insight: Pretend you’re delegating to a literal but fast intern. If it must appear, name it. If it must not appear, forbid it.

You don’t need a hundred “prompting tricks.” You need one reliable skeleton.

• Ask — the one sentence task you want done.

• Role — the voice and expertise you want applied.

• Constraints — measurable rules (length, tone, do/don’ts).

• Format — the exact output shape (bullets, steps, table, JSON).

Add a final layer—acceptance criteria—so success is testable by anyone on your team. Think of them as your mini-rubric.

⚠️ Pitfall: “Be concise” is not measurable. “≤120 words” is.

Let’s take the common disaster prompt—“Explain cloud storage”—and walk it to something the model can’t mess up.

First, name the target. Who is this for? “A non-technical adult” is clearer than “anyone.” Second, give the model a voice. Patient educator beats vague neutrality. Third, fence the language and size. No acronyms. ≤100 words. Fourth, choose the shape. Bullets, not paragraphs. Fifth, write checks you can verify in ten seconds.

Here’s the transformation:

Vague → Precise (with acceptance criteria)

Read it once and notice how your brain now knows what a correct answer looks like. That’s the whole game.

Set this at the start of a chat and reuse it. It keeps the model calm and format-focused.

Keep it in a snippet manager. Paste once. Enjoy fewer surprises.

Bullets (fast consumption). Use bullets for skimmable updates, highlights, and “what changed.” Specify how many and how long.

A: List habits that improve sleep for new parents.
R: You are a pediatric sleep coach.
C: ≤90 words total; plain language; no medical claims or products.
F: 6 bullets; each ≤12 words; last bullet is a caution.

Acceptance: bullets == 6; each ≤12 words; total ≤90; last bullet contains a caution

Numbered steps (procedures). Steps are for action. Start each line with a verb; keep one sentence per step to avoid drift.

Tables (comparisons). Tables are where models love to freestyle column names. Don’t let them.

JSON (machine-friendliness). When another tool will read the output, lock keys and counts.

💡 Insight: If the output will feed another system (slides, trackers, scripts), your Format and Acceptance are not optional—they’re your contract.

If extra paragraphs sneak in, your format was underspecified. Add “no extra sections” and restate counts. If jargon appears, make it illegal: “no acronyms; define new terms in one sentence.” If it ignores you twice, halve the scope and tighten the limits. Smaller asks get crisp results.

A reliable order of operations when rules conflict: Format → Constraints → Role. Format first keeps your structure intact; you can polish tone later.

Imagine you’re sending a one-slide update to executives. You have five bullet slots and 90 words to spend. You need status, a risk, and a next step. Here’s a prompt that ensures you get exactly that:

Run it, then check your criteria like a flight checklist. If a rule is broken, state it again, more literally, and rerun.

You’ll rewrite two vague prompts. Type your answers in your own editor so you can tweak and reuse them.

A. Event recap (bullets). Vague: “Write a quick event recap.” Your goal: 6 bullets, CTA in the last bullet, tight word budget.

One strong answer to compare with:

B. Onboarding plan (steps). Vague: “Help me onboard a new teammate.” Your goal: 5 steps, step 3 ends with a three-item checklist, no internal tool names.

One strong answer to compare with:

Score yourself with this quick rubric (0–2 each):
Ask specific to a reader (0–2). Role guides tone (0–2). Constraints measurable (0–2). Format explicit (0–2). Acceptance testable (0–2).

• Name the shape you want (bullets, steps, table) and count it.

• Make rules measurable so you can check them fast.

• Freeze voice with a role that fits your audience.

• Add acceptance criteria so anyone can verify success.

• Keep a system prompt loaded so every task starts on rails.

That’s the whole craft. Clear asks. Specific roles. Hard constraints. Exact formats. You’ll spend less time fixing outputs and more time using them.