What is prompt engineering?
Prompt engineering is designing and refining the input you give a language model so that its answer serves a real purpose. OpenAI Academy describes it in almost those words, as designing and refining your input to get the best possible answer. The input can be typed text, a pasted document, a shared file or an image.
The phrase covers two different jobs. Developers who build software on top of a model write system instructions, test them against evaluation sets and keep them under version control; the OpenAI and Anthropic developer documentation is written mainly for them. Everyone else types requests into ChatGPT, Claude, Gemini or Copilot as part of ordinary work. This guide is for the second group: marketers, account managers, recruiters, analysts and operations staff who want reliable answers, not a new career.
For that second group, prompt engineering is closer to writing a good brief for a capable colleague than to programming. Anthropic suggests thinking of the model as a brilliant but new employee who lacks context on your norms and workflows. Its test is practical: show the prompt to a colleague with little context on the task and ask whether they could follow it. If they would be confused, the model is likely to be confused too.
What goes into a good prompt?
A good prompt states the task, gives the context, supplies or points to the source material and describes the answer you want. The four major vendors use different words for this, but their advice converges on the same handful of elements.
Microsoft describes four parts for Copilot: goal, context, source and expectations, and notes that only a clear goal is strictly required. OpenAI Academy describes three steps: state the task, add helpful background or documents, and describe the output you want, including role, audience or format. Google advises Gemini users to use natural language, be clear and concise, provide context and use specific keywords. Anthropic asks for clear, explicit instructions and for the reason behind them, because knowing why a rule matters helps the model apply it.
Task
One verb and one deliverable. Draft, summarise, compare, extract, rewrite or critique, and say what the result is: a 150-word email, a table, three options.
Context
Who the reader is, what decision the output supports and what has already happened. This is the part most prompts leave out.
Source
The material the answer must rely on: pasted notes, an attached file or a named document. Say whether the tool may go beyond it.
Constraints
Length, tone, facts that must stay unchanged and things the answer must not infer, such as results, dates or names that were not supplied.
Output shape
Format, headings, a table with named columns, or a short list of questions when information is missing.
Check
What you will verify before using the answer. Writing it down keeps the review honest.
| Vendor guidance | Their terms | What it adds |
|---|---|---|
| Microsoft Copilot | Goal, context, source, expectations | Name the file, email or page to use; expect back-and-forth |
| OpenAI ChatGPT | Task, background, desired output | There is no single perfect way to prompt; refine in the conversation |
| Google Gemini | Natural language, clarity, context, keywords | Break complex work into separate prompts |
| Anthropic Claude | Clear and direct, context, examples, structure | Explain why a rule matters; use examples to steer format |
How do you write a good prompt, step by step?
Write the prompt in the order a reviewer would read it: the job first, then the facts, then the rules and the format. Here is the sequence applied to a fictional 40-person freight forwarder, Harbour Lane Logistics, whose account manager needs to tell a client about a delayed shipment.
1. Decide the deliverable
An email to a client contact, under 150 words, that explains a delay and gives the next update time.
2. Gather the facts first
The supplied facts: the container is held at the port for a customs document check; the new estimated delivery is Thursday; the next update will be sent by 4pm tomorrow. Nothing else is known.
3. Name the reader
A procurement lead who needs to tell their own warehouse team. They want dates, not reassurance.
4. Set the boundaries
Do not speculate about why the check happened, do not promise Thursday as certain and do not apologise more than once.
5. Describe the shape
Subject line, three short paragraphs, then a single line listing anything the account manager should confirm before sending.
6. Read and revise
Compare the draft with the facts sentence by sentence, then give one specific revision instruction rather than starting again.
What does a weak prompt and a strong prompt look like side by side?
A weak prompt names a topic; a strong prompt names a reader, a job, a source and a limit. Using the Harbour Lane example, the weak version is: Write an email to a client about a shipment delay. A typical response to that request is polite and confident, and it invents things: a reason for the delay, a firm new date and an offer of compensation nobody approved.
The stronger version: Draft an email from an account manager to a client procurement lead. Purpose: explain a shipment delay so they can update their warehouse team. Use only these facts: the container is held at the port for a customs document check; the estimated delivery is now Thursday; we will send the next update by 4pm tomorrow. Treat Thursday as an estimate. Do not give a reason for the check or offer compensation. Keep it under 150 words, plain and calm, with one apology at most. After the email, list any detail I should confirm before sending.
The second prompt is longer, but length is not the point. Every added sentence gives the reviewer something to check: the facts used, the claims avoided, the word limit and the questions raised. If the draft still says the delay was caused by a paperwork error at the client's end, the reviewer can see immediately that the model went beyond the source and cut that sentence.
What are some advanced prompting techniques?
The techniques that matter most beyond the basics are examples, structure, splitting large tasks, asking for options and stating instructions positively. None of them requires technical skill.
Vendors do not agree on every detail, and that disagreement is useful to know. Anthropic advises placing long documents near the top of a prompt, above the question. Microsoft says later parts of a Copilot prompt are likely to be emphasised more and suggests putting source references towards the end, and OpenAI's developer guide says supplied context is usually best near the end. The honest conclusion is that placement depends on the tool, so test both orders on a task you care about rather than adopting a universal rule.
Show an example
Paste one or two examples of the output you want, such as a past status update your manager liked. Anthropic calls examples one of the most reliable ways to steer format, tone and structure, and suggests three to five varied examples for best results when you have them.
Separate instructions from material
Use headings or labelled blocks such as Instructions, Source notes and Example, so the model can tell your rules apart from the text it is working on. Anthropic recommends tags for this; plain headings work in any chat tool.
Break the task up
Ask for an outline, approve it, then ask for the draft. Google advises splitting related tasks into separate prompts, and OpenAI Academy suggests breaking large tasks into smaller parts.
Ask for options and priorities
Request three alternatives with the trade-off of each, and say what matters most: accuracy, brevity or speed.
Say what to do, not only what to avoid
Microsoft and Anthropic both recommend phrasing instructions positively. Write in short plain paragraphs rather than do not use bullet points. Keep a do-not rule only for facts or claims that must never appear.
Invite questions
End with: before drafting, ask me up to three questions if anything essential is missing. This exposes gaps the model would otherwise fill with guesses.
Why does the first answer rarely finish the job?
The first answer is a draft to react to, and the vendors say so. Microsoft notes that the first result often is not the final or best response and that you should expect some back-and-forth. OpenAI Academy recommends refining an answer in the conversation instead of starting over.
A good follow-up names the specific problem. Make it shorter is weaker than: cut the second paragraph, which repeats the first, and move the next update time into the opening sentence. Microsoft also notes that the same prompt can produce different results each time, so a prompt that worked once is not proof it will work every time. For a recurring task, run it on several inputs before you treat it as reliable.
Want to see how this connects to a role?
Explore the planned team trial ↗Do ChatGPT, Claude, Copilot and Gemini need different prompts?
The fundamentals are the same in every assistant; what differs is what each tool can see. A clear task, real context, bounded sources and a defined format will improve results in all four.
The practical differences come from integration. Microsoft 365 Copilot can be pointed at named files, meetings, people and emails by typing a forward slash, although what you can reference depends on your licence and settings. Assistants built into an office suite, such as Gemini in Google Workspace, may reach work files depending on your plan and administrator settings. Standalone chat assistants work mostly from what you paste or upload, unless your organisation has connected other sources. Before writing a prompt that says use the latest pricing sheet, check that the tool can actually reach it; otherwise it may answer from general knowledge without saying so.
For tool-specific examples, see the cluster guides on ChatGPT prompts for work and Copilot prompts, and the wider guide to AI tools at work.
What are the most common prompt mistakes at work?
The common mistakes are asking for polish without supplying facts, relying on magic phrases and trusting a confident answer. Each one is easy to spot once you know it.
Role without substance
Act as a world-class marketer tells the model a costume, not a task. A one-line role can help with tone, but it does not replace audience, facts and limits.
No source
Asking for a market summary without supplying material invites general statements and, sometimes, invented figures.
Vague revision
Make it better gives no direction. Name the sentence, the problem and the change.
One giant prompt
Combining research, analysis, drafting and formatting in one request makes errors harder to find. Split it.
Pasting what you should not
Client contracts, personal data or unreleased figures belong only in tools your company has approved for that data.
Treating fluency as evidence
A well-written answer can still contain a false statement. Check it against the source.
How should you check what comes back?
Check every factual sentence against the source you supplied, and check anything the source did not contain twice. This is where prompt engineering and review meet: a well-scoped prompt makes the check faster because you know exactly what the answer was allowed to say.
Microsoft states plainly that Copilot responses can be inaccurate and should be reviewed and validated. That applies to every assistant. Look for added causes, dates, names, numbers and promises. Recalculate important figures independently. If the model cites a source, open it and find the passage. The guide on how to fact check AI output walks through this in detail.
How do teams learn prompt engineering?
Teams learn prompting fastest by practising on their own recurring tasks and comparing results against a clear standard. A generic course on prompt patterns helps less than repeated, reviewed attempts at the emails, summaries and analyses each person actually produces.
A useful practice loop has four parts: a realistic task with supplied facts, a first prompt, a revision, and a short note explaining what the person changed and why. A manager or peer reviews the note, not just the final text. Prompt engineering training that skips the review step tends to reward long prompts rather than good judgement.
Tosheo.ai is being built around that loop. The planned product gives each employee a short diagnostic, then 10 to 15 minute role-specific lessons that end in a hands-on task with a real AI tool, graded against a visible rubric with specific feedback. The product is pre-build, so this describes planned scope rather than results. If you want to see it applied to work tasks first, the companion guide on prompt engineering for business has more worked examples.
When is prompt engineering not the answer?
Prompt engineering cannot fix a task that needs data the tool does not have, a decision only a person can make or a tool that is wrong for the job. Anthropic's own overview makes the point that not every failure is best solved by prompting; sometimes a different model, a different tool or a different process is the fix.
If a summary keeps missing figures, the cause may be a file the assistant cannot read properly, not your wording. If a draft keeps sounding off-brand, the answer may be a real style guide and examples rather than more adjectives. If the output affects someone's pay, rights, health or legal position, a better prompt does not remove the need for a qualified reviewer.
Tosheo.ai is also not the right choice for everyone learning this. It is planned for non-technical staff in teams of 5 to 200 people. Engineers building applications on a model API will be better served by the OpenAI and Anthropic developer documentation, which cover system instructions, evaluations and version control, and by Anthropic's free interactive prompting tutorial.
Where should you go next in this guide?
Each guide in this section answers one practical question. Start with the one that matches the tool or task in front of you.
Prompt templates
Reusable templates and frameworks for summaries, emails, comparisons and rewrites, with guidance on keeping a team library current.
ChatGPT prompts for work
Ready-to-adapt ChatGPT prompts for common tasks, each with the facts to supply and the checks to run.
Copilot prompts
Prompts for Microsoft 365 Copilot in Outlook, Teams, Word and Copilot Chat, including how to point Copilot at the right file.
Prompt engineering for business
A worked example of a business prompt, a revision and a review note.
AI writing for business
How to turn a generated draft into a message you are willing to sign.
Sources and scope
OpenAI Academy, Prompting resource, accessed 16 October 2026 ↗
OpenAI API documentation, Prompt engineering, accessed 16 October 2026 ↗
Anthropic, Claude docs: Prompting best practices, accessed 16 October 2026 ↗
Anthropic, Claude docs: Prompt engineering overview, accessed 16 October 2026 ↗
Microsoft Support, Get started writing prompts in Microsoft Copilot, accessed 16 October 2026 ↗
Microsoft Support, Get better Copilot responses with great prompting, accessed 16 October 2026 ↗
Microsoft Support, Refer to specific files and more in Microsoft Copilot, accessed 16 October 2026 ↗
Google Workspace Learning Center, Tips to write prompts for Gemini, accessed 16 October 2026 ↗