How to Write Better Prompts

• The CRIT framework — Context, Role, Interrogation, Task — turns vague AI prompts into precise, business-ready outputs without manual hand-holding.
• The most common prompting mistake is writing the whole answer yourself and asking AI to “make it better.” Give it the facts and the goal, and let it build the draft.
• AI is built to agree with you, not challenge you. Declaring your own bias in the prompt is what gets you an honest, balanced answer instead of a flattering one.
• Prompting is evolving into reusable “AI skills” — standardized templates that let a team run complex, compliant processes with a single command instead of rewriting prompts from scratch every time.

Ask ten people how to get better answers out of Microsoft Copilot and you’ll get ten different guesses. Most of them are wrong — not because the tool is limited, but because the prompt is.

In a recent CIT Solutions webinar, Cole, a Power Platform Developer at CIT, broke down exactly what separates a mediocre AI response from one that’s ready to hand to your boss. As tools like Copilot become standard across the business, prompt engineering has stopped being a technical curiosity and become a core skill — whether you’re a project manager, a senior analyst, or a content marketer.

The CRIT Framework: Context, Role, Interrogation, Task

CRIT is a four-part structure for getting precise, context-aware answers out of any generative AI tool — no back-and-forth required.

C - Context (The background, constraints, and boundaries)
R - Role (The expert persona the AI should adopt)
I - Interrogation (Asking the AI to identify gaps and biases)
T - Task (The specific output and formatting required)

Context: Set the Boundaries

Context is everything the AI needs to know before it starts guessing. Drafting a project update? Context includes the project’s current status, who’s reading it, and any limits — word count, confidentiality, tone.

Role: Give It an Expert to Be

Assign a role and the AI stops giving generic advice. Instead of asking for tips, try: “Act as an expert enterprise systems analyst sitting next to me.” One sentence, and the tone, depth, and vocabulary of the response shift instantly.

Interrogation: Make It Find Its Own Blind Spots

This is the step almost nobody uses — and it’s the most valuable one. Before the AI answers, have it ask you questions first: “Identify any gaps in my context and ask me three clarifying questions to ensure the highest accuracy.” Now you and the AI are aligned before a single word of content gets produced.

Task: Say Exactly What You Want

Be explicit. Bullet points, a table, or a tight paragraph? Under 200 words or no limit? No jargon? Say so. The task is where vague requests turn into usable output.

The Mistake Almost Everyone Makes

Most people over-manage the “how” and under-explain the “what.” Write out a full email yourself and ask Copilot to “make it better,” and you’ve limited the model to editing — not thinking. Give it the key facts, the tone you want, and the objective instead, and let it find the best path there. Trust it to do the work.

There’s a second trap that’s easier to miss: context overload in long conversations. More context in a single prompt is almost always good. But in a long, multi-turn thread, the model rereads the entire history with every new message — and irrelevant detail starts diluting the focus.

The fix: when the conversation shifts to a new topic, start a fresh chat. Don’t drag six unrelated threads into one context window.

Killing Bias Before It Kills Your Output

Want an unbiased answer from an AI model? Tell it your bias first.

Large language models are built to be agreeable. Write a prompt that leans toward a conclusion, and the model will likely lean with you — not because it agrees, but because agreeing is what it’s optimized to do.

Three ways to break that pattern:

• Declare your bias. “I lean toward Option A, but I want you to challenge my thinking and present a balanced critique.”
• Ask for a devil’s advocate. “Review this proposal and identify three potential points of failure I may have overlooked.”
• Set an objective standard. Ask the model to evaluate against an industry framework, not your own preference.

Name your bias out loud, and the sycophancy has nowhere to hide.

From Prompts to Skills to Agents

Basic prompting is already becoming something bigger: reusable “AI skills” and, eventually, autonomous agents.

Individual Prompts (Manual, ad-hoc)
└──> AI Skills (Reusable, standardized templates)
└──> Autonomous Agents (End-to-end automated workflows)

Over the next one to two years, expect the shift to move from writing prompts one at a time to building and managing skills. At CIT, our development team already uses this approach to standardize processes like Copilot readiness assessments — running consistent, formatted, compliance-ready reports on a single command, instead of re-explaining the task every time.

Before rolling out Microsoft Copilot across your tenant, get your data security and permissions in order first. A standardized CRIT-built skill can analyze tenant data, flag oversharing risk, and generate the same report format every time — no matter who runs it.