The Anatomy of a Good AI Prompt

Why some prompts work better than others covered the diagnosis. This is the actual anatomy, the reusable structure underneath every strong prompt, so instead of copying someone else's prompt and hoping it transfers to your situation, you can build your own for whatever you're actually working on.

ROLE
  +
CONTEXT
  +
TASK
  +
CONSTRAINTS
  +
EXAMPLES
  +
OUTPUT FORMAT

Role

Who the AI should act as while answering. This is a system prompt's job when it's persistent across a whole conversation, covered in full here, and it's just as valuable stated once, inline, for a single prompt.

"You are a senior Microsoft Intune administrator with deep experience in Win32 app packaging and enterprise compliance policy."

Context

The specific situation the task is happening in: your environment, your constraints, why this matters right now.

"I manage 400 Windows 11 devices through Intune for a mid-sized financial services company. We're currently being audited for compliance with a client's security requirements."

Without this, the model has no way to calibrate how formal, how detailed, or how risk-averse the answer needs to be. A financial-services audit context and a small startup's "let's just get this working" context call for genuinely different advice to the same technical question. Of all six ingredients, this is the one that matters most, more than any clever wording choice elsewhere in the prompt.

Task

The specific instruction, stated as an action, not a topic. This is the exact distinction the diagnostic post covers in depth: "PowerShell scripts" is a topic. "Write a PowerShell script that checks BitLocker status on all Intune-managed devices and flags any that aren't encrypted" is a task.

"Write a PowerShell script that queries BitLocker status via the Microsoft Graph API and outputs a list of non-compliant devices."

Constraints

What the output can't assume, must respect, or must avoid. This is the ingredient most people skip entirely, and it's often the one doing the most work.

"Must use Microsoft Graph PowerShell SDK, not the legacy AzureAD module. Must not require local admin rights beyond what Intune's own remediation script context already provides. Must handle devices that haven't checked in recently without erroring out."

Examples

Showing the model what a correct input-output pair actually looks like, known as few-shot prompting, is the single most underused ingredient in this whole structure, and often the highest-leverage one for tasks with a specific format or tone.

"Here's the format I want for each flagged device: [DEVICE NAME] - [LAST CHECK-IN DATE] - [BITLOCKER STATUS] - [ACTION NEEDED] Example: LAPTOP-4471 - 2026-09-18 - Not Encrypted - Trigger remediation"

Output Format

The final shape the response should take: not just structure, but medium. A script versus a table versus a short paragraph versus a checklist. The full technique for making this actually reliable, especially for JSON and other machine-parsed formats, is covered here.

"Return only the PowerShell script in a single code block, followed by a two-sentence explanation of how it authenticates."

What Breaks When You Skip One

The Whole Thing, Assembled

That's six sentences, not six separate steps to run through. Every ingredient from this post is in there exactly once, and the result is a prompt that produces a script matching a real environment on the first try, rather than a generic starting point that needs three more rounds of clarification to get there.

Which of these six do you find yourself leaving out most often? Examples is the one I see skipped even by people who already know to include role, context, and constraints, probably because it feels like extra effort for something the model "should just know." Drop a comment with your own habit.