How do you brief AI so the first answer is close?

The idea

Brief AI the way you would brief a capable new colleague in their first week. They are quick and they work hard, but they don't know your team, your reader or which file is the right one. Anthropic's current guide puts it this way: think of Claude "as a brilliant but new employee who lacks context on your norms and workflows" (as of Oct 2026).

What a new colleague would ask you is exactly what the request needs to say. Leave it out and the AI usually does not stop to ask. You type "write an email about the delay" and leave open which delay, who it is going to and what you want them to do. A colleague would stop and ask, but the model usually just answers, so it fills the gaps with a safe, typical choice. The email is fine, and it is for nobody in particular.

Clever wording is now the smallest part of this. OpenAI's help page says that as ChatGPT grows more intuitive, "you can rely more on natural, goal-driven language and less on perfect phrasing" (as of Oct 2026). Anthropic's engineers, writing about how to set up AI agents, said "the exact formatting of prompts is likely becoming less important as models become more capable" (29 Sep 2025). Threats, tips and magic phrases tell the AI nothing about what you want. Spend those words on the brief.

What is new this autumn. The same brief now works for a quick answer and for a whole task. Since 16 Sep 2026, Claude takes both in one ordinary conversation, with no mode to choose. Since 9 Jul 2026, ChatGPT has Chat for questions and Work for finished files in one app. So a good brief is no longer only how you get a better reply. It is how you hand over the work.

The four parts of a brief

1. The goal. What you want to end up with, and what it is for. "An email that gets the project lead to confirm a new date", not "an email about the delay".
2. Who it is for, and how they will use it. Put what is in your head into words: who it's for, and how they'll use it. "My director, reading on a phone between meetings" gives it more to work with than any adjective.
3. The source to use. The file, the page or the notes it should work from, and, when it matters, "use only this". Without a source it fills the gaps with what usually comes next, and likely isn't the same as true.
4. What good looks like. The format (email, table, three bullets, a document, a five-slide deck), the length, the tone, and anything it must include or avoid. When you can't explain the style you want, show examples: Anthropic's guide calls examples one of the most reliable ways to steer the output, and suggests 3 to 5 that vary enough that it "doesn't pick up unintended patterns" (as of Oct 2026).

You do not need all four every time. Google's prompt guide for Gemini in Docs says the same of its own four areas: "You don't need to use all 4" (as of Oct 2026). Add a part whenever a new colleague would have asked about it.

One line on why helps too. Anthropic's guide says explaining the reason behind an instruction can help the model understand your goals. "Keep it under 100 words" is good. "Keep it under 100 words: they read it on a phone" is better, because the AI can now judge everything else by that.

Then improve it in one or two turns

The first answer is a draft. Treat the next message as feedback to a colleague: say exactly what to change and why.

• "Make it better" gives it nothing to go on.
• "Shorter, under 80 words, and put the new date in the first sentence" does.

OpenAI's page says to treat prompting as a conversation and refine your requests based on the first answers; Google's says to use follow-up prompts to review and refine (both as of Oct 2026). If you are on a third or fourth turn, stop and look at the brief: a part is probably missing. Add it to the first message and start again.

A worked example (made up)

Before: Write an email about the delay.

After:

Goal: an email to the client's project lead that says the report will be a week late and asks them to confirm the new date. Who it is for: they read email on their phone between meetings, so the new date has to be in the first line. Source: use the attached timeline for every date, and nothing else. What good looks like: under 120 words, plain and sorry without grovelling, and one clear question at the end.

Second turn: Good. Cut the second paragraph and keep the apology to one sentence.

That is the whole skill. The six-part brief in the Agent Playbook adds three more parts for agents (what it may touch and must not, its limits, and when to stop and ask), because an agent acts as well as writes. That is the lesson How do you hand a task to an agent?.

The ten-second test

Before you send a request that matters, show it to a colleague with little context on the task and ask them to follow it. Anthropic's guide says that if they'd be confused, Claude will be too. "The ten-second test" is our name for it; the idea is from Anthropic's guide. With no colleague nearby, read your request as if it were your first day.

Write the background once, not every time

If you type the same background into every request (your role, your team, your house style), save it once. All three assistants can hold it for you.

A skill is the same file in all three: a SKILL.md with a name, a one-line description and the steps. The lesson What is a skill? shows how to write one.

Pick the model last

The brief matters more than the model. Start with the default, move up only when the answer misses something that matters, and move down when you want speed or want to save your limit.

What you can pick depends on your plan, so check your own menu. Anthropic's quick test works in any tool: run the same brief on a lighter and a heavier model, and see whether the lighter answer misses anything.

Common mistakes

• The task with no reader. "Summarise this report." The AI does get the report, but not who it is for or how they will use it, so the summary is fine but for nobody in particular.
• Wording tricks instead of a brief. Threats, tips, "please" and magic phrases tell the AI nothing about what you want.
• No source, then trusting dates and figures it supplied from memory.
• Pasting confidential material to give it context. Check your workplace's rules first and use the tool your workplace approves (this is not legal advice).
• A single example. Anthropic's guide asks for 3 to 5 that vary, so the AI doesn't copy patterns you never meant.
• "Make it better" turns, three or four of them, instead of fixing the brief.
• Everything in one request. OpenAI suggests breaking a complex task into smaller, focused requests.
• Typing the same background every time instead of saving it once as standing instructions or a project.
• Reaching for the heaviest model first. It is slower and uses your limit faster. Fix the brief, then move the model up if the answer still misses something.

What to check before you trust it

• It used your source. Ask it to say where each date, figure or quote came from in the file you gave it, then look.
• It fits the reader and the use. Read it as the person it is for, on the screen they will use.
• Nothing was added. Treat any name, number, promise or date that is not in your source as a guess until you have checked it.
• Hedges survived. If your source says "about" or "in one survey", the answer should too.
• Before it goes out, run the checks in the next lesson.

As of: product steps as they were on the vendors' own pages on 16 Oct 2026.

• Neosaipien's platform reference (checked on the vendors' own pages, 16 Oct 2026): model names, plans, work modes, projects, instructions, skills and the comparison
• Anthropic, Prompting best practices
• Anthropic, Effective context engineering for AI agents (29 Sep 2025)
• Anthropic, Claude Cowork and chat are now one Claude (16 Sep 2026)
• Anthropic, Claude Cowork and chat are one Claude (help page)
• Anthropic, Build live dashboards and animate explainers with Claude (8 Oct 2026; Docs, Slides and Design on every plan)
• Anthropic, Change the model, effort, and thinking settings
• Anthropic, Choosing the right Claude model (Claude Academy)
• Anthropic, Understanding Claude's personalization features
• Anthropic, What are projects?
• OpenAI, How do I create a good prompt for an AI model?
• OpenAI, ChatGPT Work and Codex
• OpenAI, GPT-6 and other models in ChatGPT
• OpenAI, ChatGPT release notes (9 Jul 2026, ChatGPT Work; 10 Sep 2026, files from Library)
• OpenAI, Projects in ChatGPT
• OpenAI, ChatGPT custom instructions
• Google, Learn about prompts for Gemini in Docs, Sheets, Slides, Vids & Forms
• Google, Get started with Google Workspace with Gemini
• Google, About Gemini models
• Google, Upload & analyze files in Gemini Apps (Add Files)
• Google, Use apps connected to Gemini with a work or school account
• Google, Gemini Apps release notes (13 Oct 2026, the Customize menu)
• Google, NotebookLM is now Gemini Notebook (16 Jul 2026)
• Google, Manage your tasks and workflows in Gemini Apps
• Agent Skills, overview (one SKILL.md format)