# Brief an AI agent

A free starter skill from Neosaipien. Version 1.1, as of 16 October 2026.

You type a short, loose request, such as "chase the overdue invoices". The skill turns it into a clear brief you can paste into an AI agent, so the agent knows what to do, what it must leave alone, and when to stop and ask you.

## What it does

For each request you give it, you get back:

- **A paste-ready brief in six parts:** the goal; what done looks like; the sources the agent may use; what it may touch and what it must not touch; when to stop and ask you; and how you will check the result.
- **The risks to watch:** what private data the agent could see, and which of its actions would be hard to undo.
- **Up to three questions,** only where something is truly unclear. Where it does not know a fact (a budget, a date, a folder), it says so or leaves a marked blank. It never makes one up.

It starts from safe defaults: **draft only, ask before sending, never delete.** If you give it your team's rules, it carries them into every brief, by rule number.

It writes the brief. It does not do the task.

## Who it is for

Managers and staff who are starting to hand work to an AI agent: an ordinary Claude conversation (Claude now takes on tasks in any chat), ChatGPT Work, or a task in Gemini.

## Why it matters now

All three tools now take a goal and act on it, and they decide for you less than you might think (as of Oct 2026):

- **Claude:** since 16 Sep 2026, Cowork is part of ordinary Claude, so any conversation can take on a task. By default it asks before taking an action, but you can let it keep working and check in only when something needs a closer look.
- **ChatGPT:** Work asks you to "describe the outcome you want" with your rules and how you will judge the result, and to approve important actions. Its app permissions run from "Always ask" to "Allow all actions".
- **Gemini:** a task works through your connected apps, skills and a browser. It confirms some actions and sign-ins.

So the brief, not only the tool, should say what needs your yes. The UK's National Cyber Security Centre (20 Aug 2026) says to decide what an agent is to achieve, what it may do, and "when it should stop and seek human approval", and warns not to rely on instructions alone.

## What is in the download

| File | What it is |
|---|---|
| `brief-an-agent.zip` | The skill, ready to upload to Claude. Inside: `brief-an-agent/SKILL.md` (the instructions) and `LICENSE.txt` |
| `skill/brief-an-agent/` | The same skill as an open folder. `SKILL.md` is the file for ChatGPT and Gemini |
| `portable/paste-in.txt` | A plain-text copy for a plan without skills: paste it into a project's instructions, or at the start of a chat |

## Add it to your AI tool (as of Oct 2026)

All three tools now take skills in the same open format, a SKILL.md file. The format began at Anthropic and is now an open standard. So this is one skill, not a version per tool. Use the tool your company approves for work data.

If you cannot see a menu below, the usual reasons are a rollout still in progress, your plan, your country or an admin setting.

### Claude

**Plans:** Pro, Max, Team and Enterprise, says Anthropic's skills page (its pricing table also ticks Free).

1. In **Settings > Capabilities**, turn on **Code execution and file creation**. Skills need it.
2. Go to **Customize > Skills**.
3. Upload `brief-an-agent.zip`. It holds the SKILL.md file.
4. To use it, describe the task in a chat, or type **/** and pick **brief-an-agent**.

Anthropic does not review skills made by other people, so read a skill's files before you turn it on. This one is plain text: open `skill/brief-an-agent/SKILL.md` to read it. A skill you add in Claude also follows you into Claude Code when you sign in with the same account.

### ChatGPT

**Plans:** skills are a beta on Free and Plus, and generally available on Enterprise and Edu (since 9 Jul 2026). ChatGPT keeps skills inside plugins, in the Plugin Directory.

1. Start a new chat and type **@skill-creator**.
2. Attach `skill/brief-an-agent/SKILL.md` and write: "Make a skill from this file. Keep its name, description and instructions exactly as written."
3. Check that the skill it shows you matches the file.
4. To use it, describe the task, or type **@** and pick **brief-an-agent**.

### Gemini

**Plans:** 18 or over, with a personal Google Account. Google says work and school accounts get skills "soon".

1. Open the left menu, click **Customize**, then **Skills**.
2. Make a new skill by uploading a SKILL.md: choose `skill/brief-an-agent/SKILL.md`.
3. To use it, describe the task, or type **/** (Google says **@** is coming) and pick **brief-an-agent**.

Gemini wants skill names in lowercase with hyphens, which this one has, and allows up to 100 active skills.

### No skills on your plan? Use a project

- **Claude or ChatGPT:** make a project for this job and paste the whole of `portable/paste-in.txt` into the project's instructions. Every chat in that project then follows it. In ChatGPT: select **New project** in the sidebar, name it, then add instructions from the **•••** menu under **Project settings**.
- **Gemini with a work account, or any single chat:** paste `portable/paste-in.txt` as your first message. It lasts for that one chat.

There is no custom GPT or Gem version: OpenAI plans to retire custom GPTs in favour of plugins, and Google is moving Gems to skills from Nov 2026 for personal accounts and in 2027 for work and school accounts.

## Which tool, mode and model to use

**My recommendation: Claude first, ChatGPT second, Gemini third.** (Samson, Neosaipien.) Writing the brief is a plain chat job. Running it is agent work, and that is where the three tools differ most.

| | Claude | ChatGPT | Gemini |
|---|---|---|---|
| Write the brief | Any ordinary Claude conversation, Sonnet 5.5 at its Default effort | Chat, GPT-6 at Medium | Chat, Flash |
| Where the agent runs it | The same ordinary Claude conversation; the desktop app if it needs folders on your computer | Work | Task (personal accounts on Google AI Pro or Ultra only) |
| Model for the agent | Opus 5.5 | Work with Default, or GPT-6.1 Sol | Flash or Argon |
| Keeps working when your laptop is closed | Yes | Yes (Work in the cloud) | Yes (remote browser) |
| Asks before acting | Yes, by default | Approves important actions | Confirms some actions and sign-ins |

- **Plans:** Opus 5.5 needs Claude Pro or above. ChatGPT Work is limited on Free (desktop app only), and on desktop, web and phone from Plus up. Gemini's Task needs a personal Google Account, 18 or over, Google AI Pro or Ultra, Keep Activity on and Personal Intelligence on, and is not offered in the European Economic Area, Nigeria, Switzerland or the UK.
- **With a Gemini work account:** there is no Task agent yet. Use Gemini inside Gmail, Docs, Sheets and Drive for the steps, and do the actions yourself.
- **Where another tool wins:** a job that must start each time an email or a Slack message arrives fits ChatGPT Work's Scheduled Tasks best.

## Where each tool's approval setting is (as of Oct 2026)

The brief asks the agent to stop and ask you before it sends, books, pays or moves things. The tool's own setting decides whether it really has to. Set it yourself before you hand over a brief.

- **Claude:** in the conversation, Claude asks before taking an action by default. Keep it that way for these briefs, rather than letting it keep working on its own. For each connector, a tool can be set to "Always allow", "Needs approval" or "Blocked"; choose "Needs approval". When Claude asks to use a tool in a chat, choose "Allow once", not "Always allow".
- **ChatGPT:** Work asks you to approve important actions. For connected apps, the permission choices are "Always ask", "Allow read actions", "Allow low-risk actions" and "Allow all actions". Choose "Always ask" or "Allow read actions" for these briefs, and never "Allow all actions". Its cloud browser pauses for your sign-in or confirmation.
- **Gemini:** a task asks you to take over for sign-ins, and some actions need your confirmation. In connected apps, write actions always need your confirmation. Google advises not to type passwords or payment details into the task.

In all three, type passwords into the website yourself, never into the chat.

## How to use it

Start a new chat and ask in plain words, for example:

- "Brief an agent to chase the overdue invoices before month end."
- "I want to hand these three jobs to Claude. Here are our team rules." (then paste or attach them)
- "Turn this into a task for the agent: clean up the shared project folder before the new starter arrives."

Give it your team's rules if you have them, even a few lines. Answer its questions, fill in any [TO DECIDE] blanks, then copy each brief into your agent: an ordinary Claude conversation, ChatGPT Work, or a Gemini task.

## What you must check

The skill ends every answer with these four checks, because they are things it cannot know or control:

1. **The "must not touch" part matches your team's real rules and data.** Add anything the skill could not know.
2. **Set the approval setting in your agent tool yourself,** so it asks before it acts. The brief cannot change that setting, and each tool offers options such as "Always allow" or "Allow all actions". The section above says where the setting is in each tool, as of Oct 2026.
3. **No password, customer record or private detail went into the brief.** The skill leaves these out if you paste them, and tells you, but check.
4. **Check the agent's work against the "how I will check it" part** before you rely on it or send anything.

## What it does not do

- It does not carry out the task, send anything, open links, connect to your apps or change any setting.
- It does not change, or know, the approval settings in your agent tool.
- It does not guess facts it was not given, and it does not decide about people (hiring, pay, performance).
- It does not make an agent safe on its own. Instructions help, but the tool's settings, your team's rules and your own check still matter. There is no promise of any result.

## Your data

The skill is one text file of instructions. It has no scripts and no links, needs no internet access, and sends nothing anywhere. It works only with what you type or attach in your assistant, and your assistant's own data terms apply.

## Compatibility, as of October 2026

| Tool | Who can add skills (vendor's own pages) | Which file |
|---|---|---|
| Claude | Pro, Max, Team and Enterprise (the pricing table also ticks Free), with code execution on | `brief-an-agent.zip` |
| ChatGPT | A beta on Free and Plus; generally available on Enterprise and Edu | `skill/brief-an-agent/SKILL.md`, through @skill-creator |
| Gemini | 18 or over, personal Google Accounts; work and school accounts "soon" | `skill/brief-an-agent/SKILL.md` |
| A plan without skills | Everyone, through a project's instructions or one chat | `portable/paste-in.txt` |

"Who can add skills" is what each vendor's own page says, as of October 2026. These tools change often.

Built to the open Agent Skills format and checked against each tool's published requirements on 16 October 2026. Tools change every month, so these steps carry a date.

## Licence and contact

Free, under the Neosaipien free licence, Version 2. Use it as much as you like, at work or at home, in any organisation. Change it to fit your team, and hand the changed version to people inside your own organisation, team, class or group, with the credit, the date and a line saying it was changed. Share unchanged copies with anyone, free of charge. Do not sell it, and do not publish a changed version outside your own group. Full terms in `LICENSE.txt`. No warranty: you check every brief and everything your agent does.

Neosaipien is not affiliated with Anthropic, OpenAI or Google. Product names belong to their owners.

Questions, or help writing skills in your own team's rules: samson@neosaipien.com, neosaipien.com

## Sources for the setup steps (all read 16 October 2026)

- Anthropic, Skills overview: https://claude.com/docs/skills/overview.md
- Anthropic, Claude plans and pricing: https://claude.com/pricing
- Anthropic, What are projects: https://support.claude.com/en/articles/9517075-what-are-projects
- Anthropic, Connectors, skills, and plugins: https://claude.com/docs/extend/overview
- Anthropic, Change the model, effort, and thinking settings: https://support.claude.com/en/articles/8664678-change-the-model-effort-and-thinking-settings
- OpenAI, Skills and plugins: https://learn.chatgpt.com/docs/skills-and-plugins.md
- OpenAI, Build skills: https://developers.openai.com/plugins/build/skills.md
- OpenAI, ChatGPT release notes: https://help.openai.com/en/articles/6825453-chatgpt-release-notes
- OpenAI, ChatGPT plans and pricing: https://chatgpt.com/pricing/
- OpenAI, Projects in ChatGPT: https://help.openai.com/en/articles/10169521-projects-in-chatgpt
- OpenAI, GPT-6 and other models in ChatGPT: https://help.openai.com/en/articles/20001354-gpt-6-and-other-models-in-chatgpt
- Google, Create and manage skills for Gemini Apps: https://support.google.com/gemini/answer/17094296
- Google, About the transition from Gems to skills: https://support.google.com/gemini/answer/18560919
- Google, About Gemini models: https://support.google.com/gemini/answer/18648722
- Agent Skills overview: https://agentskills.io/

## Sources for the agent and approval sections (all read 16 October 2026)

- Anthropic, Claude Cowork and chat are now one Claude (16 Sep 2026): https://claude.com/resources/articles/cowork-is-now-claude
- Anthropic, Get started with connectors: https://claude.com/docs/connectors/overview
- OpenAI, ChatGPT Work and Codex: https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex
- OpenAI, Connected apps in ChatGPT: https://help.openai.com/en/articles/11487775-apps-in-chatgpt
- OpenAI, Using cloud browser in ChatGPT: https://help.openai.com/en/articles/20001280-using-cloud-browser-in-chatgpt
- Google, Manage your tasks and workflows in Gemini Apps: https://support.google.com/gemini/answer/17094507
- Google, Connect and manage custom apps for Gemini Apps: https://support.google.com/gemini/answer/17209137
- UK NCSC, Managing the cyber risk of agentic AI (20 Aug 2026): https://www.ncsc.gov.uk/blogs/managing-the-cyber-risk-of-agentic-ai