The Anatomy of a Prompt That Actually Works

Every prompt that consistently works is built from the same six parts: a role, context, the task, a format, constraints, and examples. Miss one and you get the familiar failure modes. Generic output means no role. Confident nonsense means no context. A wall of text means no format. Once you can see the bones, you can diagnose any bad output in seconds.

What are the six parts of a prompt that works?

1. Role: who the AI should be

"You are a B2B demand gen lead reviewing a landing page" produces different output than a bare request. The role sets vocabulary, judgment, and what the model treats as obvious. One sentence is enough.

2. Context: what it cannot guess

The model does not know your audience, your product, your goal, or what you already tried. This is the part people skip most, and it is the part that matters most. Two or three sentences of situation beat two paragraphs of instructions.

3. Task: one specific job

"Write a follow-up email to a prospect who went quiet after a demo" beats "help me with sales emails." If the task has two jobs in it, split it into two prompts.

4. Format: what the output looks like

Say whether you want a table, five bullet points, a 120-word email, or JSON. If you do not pick the shape, the model picks for you, and it loves headers and long lists.

5. Constraints: the guardrails

Length limits, tone, words to avoid, things that must be included, things that are off the table. Constraints are what make the second draft unnecessary.

6. Examples: show, do not describe

One example of what good looks like outperforms a paragraph describing it. Paste a past email you liked, a table in the right shape, or a before and after.

What does weak vs strong actually look like?

Weak: "Write a cold email for my agency."

Strong: "You are an SDR at a HubSpot implementation agency (role). We help 20-200 person B2B companies clean up their CRM and launch AI agents; the prospect is a RevOps manager whose team just downloaded our audit checklist (context). Write a three-sentence cold email referencing the download and offering a 15-minute teardown (task). Plain text, no subject line yet, under 80 words (format). No buzzwords, no exclamation points, one specific question at the end (constraints). Here is an email in our voice: ... (example)."

Same model, same day, wildly different output. The difference is not talent. It is parts coverage.

Do you have to assemble this yourself every time?

No, and that is the honest pitch. Knowing the anatomy makes you better at judging output, but assembling role, context, format, and constraints by hand for every task is exactly the kind of work software should do. Promptbuildr does it for free: describe the task in plain language, pick your role and your AI tool, and it generates the complete prompt with instructions for using it. It also tunes structure per tool, since the same prompt does not work in every AI. Claude responds best to XML-style structure, ChatGPT to clean labeled sections, and if you want to try Claude itself, this link gets you a free week.

One guardrail we will keep repeating: a generated prompt gets you a strong first draft fast, not a perfect answer every time. The six parts raise the floor dramatically. Your judgment still sets the ceiling. If your outputs are still off after a complete prompt, the fix is usually in context or examples, which we broke down in why your AI keeps getting it wrong.

Frequently asked questions

How do you write a good AI prompt?

Give the AI six things: a role to play, the context it cannot guess, a specific task, the output format you want, the constraints that matter, and at least one example. Most bad outputs trace back to one of those six being missing.

What is the most important part of an AI prompt?

Context. The model only knows what you tell it. Audience, goal, and what you have already tried change the answer more than any clever wording does.

How long should an AI prompt be?

As long as it needs to be to carry the six parts, and no longer. A strong prompt for a real work task usually runs 100 to 300 words. One-line prompts work only for one-line jobs.

Do I need to learn prompt engineering to get good results?

No. You need to know what a complete prompt contains, and tools can assemble it for you. Promptbuildr generates a tool-tuned prompt with instructions from a plain-language description of your task, free.