The short answer
A prompt is the text, and any files or images, you give an AI model to tell it what to do. Google's Machine Learning Glossary defines it as any text entered as input to a large language model to condition the model to behave in a certain way. A good prompt states the task, the context, the audience, the format and any examples. In 2026, ChatGPT, Gemini and Microsoft Copilot also let you save reusable prompts as skills.
What a prompt is, in plain English
A prompt is your side of the conversation with an AI model. It can be one line ("summarise this"), a full page of instructions, a pasted document, a screenshot or a combination of all of them. The model reads the whole prompt and generates the answer it predicts fits best.
Google's Machine Learning Glossary defines a prompt as any text entered as input to a large language model to condition the model to behave in a certain way. The same glossary defines prompt engineering as the art of creating prompts that elicit the desired responses. You do not need the job title to use the skill: it mostly means being specific.
Two related terms are worth knowing. A few-shot prompt includes more than one example of the answer you want, which Google's glossary describes as a prompt containing a few examples demonstrating how the model should respond. A system prompt is a standing instruction set by the app or by you that applies to every message, such as custom instructions in ChatGPT.
Why prompts matter at work
The model has no idea who you are, who will read the result or what "good" looks like in your team unless you tell it. A vague prompt gets an average answer for an average reader. A specific prompt gets something you can send.
Prompts also control cost and time. Every word you send is counted in tokens and fills part of the model's context window. Pasting three unrelated documents "just in case" can make answers worse, not better. Anthropic's documentation notes that accuracy and recall degrade as the amount of context grows.
The five parts of a prompt that works
- Task. One clear verb and output: "Write a 150-word follow-up email", not "help with this email".
- Context. What the model cannot know: the situation, what happened before, the goal.
- Audience and tone. Who reads it and how it should sound: "for a finance director who prefers short, direct notes".
- Format. Length, structure and layout: bullet points, a table, three options, plain text you can paste.
- Examples or constraints. A sample of what good looks like, and what to avoid: "keep all figures exactly as given", "say if information is missing".
Anthropic recommends one more line for factual work: explicitly give the model permission to say it does not know. Its guide on reducing hallucinations says this alone can drastically reduce false information.
Try it: the same request, before and after
Open any assistant. First, type the weak version:
Before: "Write an email about the project delay."
You will get a generic, slightly formal email that could be about anything. Now try the strong version:
After: "Write an email to our client, Dana, a marketing director, explaining that the website launch moves from 12 November to 26 November because their product photos arrived late. Tone: friendly, direct, no grovelling. Under 120 words. Offer two times for a 15-minute call next week. End with one line on what we need from her by Friday."
Compare the two. The second takes 30 seconds longer to write and usually needs no rewrite. That trade is the whole idea of prompting. For a full method, read how to write a good prompt, and for ready-made examples see prompts for emails and ChatGPT prompts for marketing.
Saved prompts: skills, projects and custom instructions
If you type the same instructions every week, save them. The main assistants changed how they do this in 2026.
- Gemini: Gems are becoming skills, reusable instruction sets that Gemini can run automatically when they match a prompt. For personal accounts aged 18+, Gems turn read-only on 17 November 2026 and are removed on 17 December 2026, converted to skills automatically.
- ChatGPT: custom GPTs are moving to plugins, which bundle skills, apps and connections. OpenAI's migration FAQ gives 11 December 2026 as the retirement date, with migration under My GPTs, Migrate to plugin.
- Microsoft Copilot: skills, reusable instructions written in natural language, are expanding from Copilot Cowork to Copilot Chat in November 2026.
- Claude: projects keep instructions and files together for a piece of work, on every plan (free accounts can create up to five). See how to set up Claude projects.
Common misconceptions about prompts
- "There is a magic phrase." Tricks like "you are a world-class expert" matter far less than clear context and a clear format. Specific beats clever.
- "Longer is always better." Longer is better only when the extra words are relevant. Irrelevant pasted material can distract the model.
- "A prompt that works in ChatGPT works the same everywhere." Different models have different habits, and even different plans of the same tool use different models. Test important prompts in the tool your team actually uses.
- "Prompting is a one-shot thing." The best results usually come from a short back-and-forth: ask, read, then say exactly what to change.
- "What I paste is private." Prompts are sent to the provider. On many personal plans they can be used for training unless you opt out, so read how to keep your data private in AI tools before pasting client data.
Related terms
- System prompt means standing instructions that apply to every message.
- Custom GPT is a saved assistant with its own prompt and files, now moving to plugins.
- Prompt injection means hidden instructions in content that try to hijack the model.
- Temperature is the setting that makes answers more or less varied.
- Token and context window set how much prompt the model can take.
Prompts to copy
The five-part prompt template
Task: [what you want, with length]. Context: [the situation and the goal]. Audience: [who reads it and what they care about]. Format: [bullets, table, email, number of options]. Rules: keep all names, dates and figures exactly as given; if something I need is missing, ask me before writing.
You get: A first draft that fits your situation, or a short list of questions if key details are missing.
Let the AI improve your prompt
Here is a prompt I plan to use: [paste prompt]. Ask me up to five questions that would make the result more useful, then rewrite the prompt using my answers.
You get: A sharper version of your own prompt, built from the details the model needed.
Turn a good chat into a reusable instruction
Look at this conversation. Write a reusable instruction I can save as a skill or custom instruction so that next time I get this result in one step. Include the audience, tone, format and rules we settled on.
You get: A saved-prompt text you can paste into a skill, project or custom instructions.