Most disappointing AI answers come from short, vague requests. "Write a follow-up email" gets a generic email. "You are writing for our client services team; draft a follow-up to this client who missed the deadline in the thread below; keep it under 120 words, friendly, and ask for a new date" gets something close to sendable. The second request is what a good manager does when delegating: say who, what, why, with what, and what done looks like. No technical skill is involved.
This guide is the prompting module from our workshops, written down. It works on ChatGPT, Microsoft 365 Copilot, Gemini and Claude, and it is meant for people who will never write code.
What is prompt engineering for employees?
Prompt engineering for employees is the practice of writing clear, complete instructions for a generative AI tool so it produces usable work, and then checking that work before relying on it. For developers the term covers system prompts and testing at scale; for everyone else it means writing a good request and knowing what to do with the answer.
The vendors agree on the basics, even if they name them differently. Google's guidance for Gemini in Workspace lists persona, task, context and format. Microsoft's guidance for Copilot lists goal, context, source and expectations, and notes that only the goal is required. Anthropic's guidance for Claude suggests treating the model as a brilliant but new employee who lacks context on your norms and workflows, and recommends three to five examples for best results. The five-part method below combines them.
What are the five parts of a good prompt?
A good prompt gives the AI a role, a specific task, the context and source material, the format you want back, and, when it matters, an example. You will not need all five every time; the more the output matters, the more you include.
| Part | What to write | Weak | Better |
|---|---|---|---|
| Role | Who the AI should act as, and who the output is for | (none) | "You are an operations coordinator writing for new warehouse staff." |
| Task | One clear action verb and the deliverable | "Help with this SOP." | "Turn the notes below into a step-by-step SOP." |
| Context | Why you need it, constraints, and the source to use | (none) | "Use only the notes below. Staff read this on a phone during a shift." |
| Format | Length, structure, tone, what to leave out | (none) | "Numbered steps, under 200 words, plain words, no introduction." |
| Examples | A sample of good output, or a past document to match | (none) | "Match the style of the receiving SOP attached." |
Two habits make the parts work better. Say what to do when information is missing ("if a step is unclear, list it as a question at the end instead of guessing"), which reduces invented detail. And state the reason behind a rule; Anthropic notes that explaining why an instruction matters helps the model deliver more targeted responses.
What does a complete prompt look like?
Here is a full prompt for a common office task, using fictional details. It takes about a minute to write and usually saves several rounds of back-and-forth.
ROLE You are a client services lead at a Chicago insurance agency.
TASK Draft a renewal reminder email to the client below.
CONTEXT Policy: commercial auto, renews Dec 15. Premium up 6% after
a claim in March. Client prefers short emails. Use only the
facts here; do not mention coverage details not listed.
FORMAT Under 120 words. Subject line first. Friendly, direct.
End by offering two call times next week.
EXAMPLE Match the tone of this past email: "Hi Dana, quick note
before your renewal..."
IF UNSURE List questions for me after the draft instead of guessing.Notice what it does not include: the client's real policy number, address or claim details. Use the minimum real information the task needs, and only in tools your company has approved. Our AI acceptable use policy template sets out what never goes in.
How should employees check AI output?
Check every output against its source before you send, file or act on it. AI tools can produce confident, fluent text that is wrong, and fluency makes errors harder to spot.
The U.S. National Institute of Standards and Technology calls this risk confabulation: the production of confidently stated but erroneous or false content. A field experiment with 758 Boston Consulting Group consultants showed why it matters: on tasks inside the AI's capabilities, consultants using it did more work, faster and to a higher standard, but on a task outside those capabilities they were 19 percentage points less likely to reach the right answer than colleagues without AI. The tool does not warn you when you have crossed that line.
A checking routine that takes a minute or two:
- Names, numbers, dates. Compare each one against the source document, not your memory.
- Anything added. Look for facts, quotes, statistics or commitments that were not in the material you gave it.
- Sources and citations. If it cites a law, article or case, open it. If you cannot find it, it does not go in.
- Math. Recalculate totals and percentages, or ask the tool to show its working in a spreadsheet.
- What is missing. Ask whether anything important in the source was left out.
- Tone and audience. Read it as the recipient would.
- Would you sign it? If your name goes on it, you own it. AI drafts; a person decides.
You can also ask the tool to help: "List every factual claim in your answer and quote the sentence in the source that supports it." It will not catch everything, but it makes gaps visible.
Which prompt patterns work for each role?
Most office work falls into a handful of patterns: summarize, draft, transform, compare, analyze and critique. Each role leans on a few of them. These are starting points we use in Day 2 labs; adapt the wording to your work and use practice data while you learn.
| Role | Pattern | Example prompt |
|---|---|---|
| Sales | Draft from a brief and a past example | "Using the call notes below and our proposal template, draft the scope section. Flag any price or date you had to assume." |
| Customer service | Draft from policy | "Using only our return policy below, reply to this customer. If the policy does not cover their case, say so and suggest escalation." |
| Operations | Transform notes into a procedure | "Turn these rough notes into a numbered SOP for a new hire, then list any step that seems missing." |
| Finance | Analyze and explain | "Here is a practice version of our monthly expense export. Explain the three biggest changes from last month in plain sentences and show the formulas you used." |
| HR | Draft and check for clarity | "Rewrite this policy paragraph at an eighth-grade reading level without changing its meaning. List anything you were unsure about." |
| Marketing | Generate options against criteria | "Give five subject lines for this newsletter under 50 characters, each in our voice from the examples below, and say which audience each suits." |
| Managers | Summarize to decisions | "From this meeting transcript, list decisions made, open questions and action items with owners and dates. Mark anything ambiguous." |
| Leadership | Critique as a thinking partner | "Here is my plan for next quarter. Argue against it as a skeptical CFO would, then list the three questions I should answer before presenting it." |
Note the line most of these share: flag assumptions, say when the source does not cover something, or list what you were unsure about. That one sentence does more for accuracy than any clever wording.
What should you do when the first answer is wrong?
Tell the tool exactly what to change instead of starting over or rewriting it yourself. Treat the first answer as a draft and give feedback the way you would to a colleague.
- Be specific: "shorter" is weaker than "cut to 80 words and remove the second paragraph."
- Point to the problem: "the date in line two is wrong; the deadline is the 15th."
- Add the missing context you realize it needed.
- Ask it to ask you: "What information would you need to do this better?"
- If a long conversation drifts, start a fresh chat with an improved prompt that includes what you learned.
Microsoft's own guidance for Copilot says to expect some back-and-forth to get the results you want. Plan for two or three rounds on anything important.
When should a prompt become a saved instruction?
Save a prompt once it has worked well on a few real inputs and the task repeats weekly or more. Saved instructions are where most of the lasting value of training ends up, because nobody has to remember the perfect wording.
Each tool has a place for them. ChatGPT and Claude both offer Projects, which keep instructions and reference files together for a recurring task. Microsoft 365 Copilot lets you save prompts and share them with a team. Google is moving Gemini's custom Gems to a new feature it calls skills. Whatever the tool, write each saved instruction the same way: what it is for, the role, task, context and format, the source it should use, and the checks a person must do before using the output. Give each one an owner and test it before the team relies on it. Our guide to building your first AI workflow takes this further.
What should never go into a prompt?
Never paste passwords, access keys, full card or bank numbers, or information your company has classified as off-limits for AI. Use work accounts on approved tools for anything involving company information.
Beyond that, the rules depend on your plan and your policy. Business plans from OpenAI, Microsoft, Google and Anthropic state that they do not train on customer data by default, while some personal plans share data unless the user opts out. That is why we recommend work accounts for work, and why a written policy matters more than any single tip. If you have not chosen a tool yet, our ChatGPT vs Copilot vs Gemini vs Claude comparison covers data controls for each.
How can employees learn prompting in Chicago?
The fastest way is practice on your own work with someone watching over your shoulder. 312 IT Consulting runs hands-on sessions on-site across Chicago and the suburbs and live online, on your team's own accounts. Day 1 covers how AI works, prompting, checking output and privacy; Day 2 builds saved instructions and labs on your own processes with practice data. A 2–3 hour AI workshop in Chicago for up to about 15 people typically starts from about $2,500, and every attendee receives a certificate of completion. Results from past sessions are on our AI training results page, and longer programs are quoted in writing after a free first call.