Anatomy of a Great Prompt: The 6 Building Blocks

Most bad prompts aren't bad because of clumsy wording. They're bad because they're missing a part — the model was never told the format, or the audience, or what good looks like, so it filled the gap with a guess. Once you can see a prompt as a small set of building blocks, writing a good one becomes less like magic and more like a checklist: is each block present, and is each one clear?

Here are the six blocks. Not every prompt needs all six — a quick question needs one or two — but when a prompt underperforms, the fix is almost always a missing block.

1. The task

The one non-negotiable block: what do you actually want done? State it plainly and specifically, and put it near the front. Modern models respond to direct instruction and stumble on hints.

Vague: Help me with my resume.
Clear: Rewrite the work-experience section of my resume to emphasize measurable impact.

If you strip a prompt down to nothing else, keep this — and make it specific. Help with X offloads the hardest decision (what kind of help) back onto the model.

2. Context

The task tells the model what; context gives it what it needs to know. This is the block people skip most, and it's the highest-leverage one — a capable model with the right context beats a clever prompt with none.

Context is the facts, background, and source material the model can't infer: who you are, what you're working with, the actual data. Paste the resume. Include the job description. Give the numbers. The model can't reason about information it doesn't have, and it won't ask — it'll guess.

3. The why — motivation

Telling the model why the task matters lets it make good judgment calls in the gaps your instructions didn't cover. It's the difference between a rule and an understanding.

Rule only: Keep it under 150 words.
Rule + why: Keep it under 150 words — it's going in a LinkedIn post and long posts get skipped.

With the reason attached, the model generalizes correctly to decisions you never spelled out. It knows why short matters, so it makes other post-appropriate choices too.

4. Constraints and guidance

The guardrails: what to include, what to prioritize, what to avoid. Every constraint you add removes a whole category of wrong answers.

A key rule here from modern prompting practice: tell the model what TO do, not what NOT to do. Write in flowing prose points at one target; don't use bullet points leaves infinite room. Positive instructions beat prohibitions almost every time.

Prioritize achievements with numbers. Use active verbs. Match the seniority level in the job description.

5. Examples

Often the fastest way to communicate what you want is to show it. One good example of the output you're after can do more than a paragraph describing it — and today's models pay very close attention to the details in examples, so make sure yours model exactly the behavior you want.

A word of caution: examples are powerful enough to backfire. On reasoning-heavy tasks, examples can bias the model toward copying surface patterns instead of thinking the problem through — so examples aren't always the right move. When to include them (and when to skip them) is its own topic, covered in Few-Shot vs Zero-Shot.

6. Output format

Tell the model the shape you want the answer in, or you'll get its default — usually a chatty paragraph when you wanted a table, or a wall of text when you wanted three bullets.

Return three versions, each as a single paragraph, labeled A, B, and C.

If you're feeding the output into software rather than reading it yourself, don't just ask for JSON — constrain it. That's a reliability topic of its own: Getting Reliable Structured Output.

Putting it together

Here's the same request with the blocks stacked — task, context, why, constraints, format:

Rewrite the work-experience section of my resume to emphasize measurable impact. (task)
Here's the current section and the job description I'm targeting: [pasted]. (context)
This is for a senior PM role, and the hiring manager skims dozens of these, so impact has to jump out fast. (why)
Prioritize bullets with concrete numbers, lead each with an active verb, and match the seniority language in the job description. (constraints)
Return the rewritten bullets as a plain list, ready to paste. (format)

Notice there's no clever phrasing anywhere in it. It works because every block a model needs is present and unambiguous — which is the whole game.

The takeaway

When a prompt disappoints, don't reach for fancier wording. Run the checklist: Task. Context. Why. Constraints. Examples. Format. Nine times out of ten a block is missing, and adding it fixes the output faster than any amount of rephrasing.

This is the working core of The Complete Guide to Prompt Engineering in 2026 — start there for the bigger picture. And if you'd rather build on prompts that already have every block in place, the prompt library is full of them.