Reusable AI setups compared: projects, skills and agents

Check the kind of setup you need

A project organises context; a reusable instruction guides a recurring task; an agent may also act in connected systems. These can overlap, but they do not provide identical permissions, sharing or memory. Google is changing its reusable-instruction options: follow the Gems and skills lesson before choosing a new team setup. Keep the task brief portable and test access with the people who will use it.

One idea, four badges

Somewhere in the four tools you've learned, you have probably found yourself pasting the same background into a fresh chat for the tenth time: "You're helping Fernway's operations team, write in plain UK English, never invent figures." A custom assistant saves that standing set-up once so you skip straight to the request. Every major tool now offers this, and underneath they are the same idea: a saved prompt with a name, some fixed instructions, and optionally a few reference documents it always draws on.

The names differ, which makes four features look like four skills to learn. They aren't. Learn the shape once:

• ChatGPT offers Projects, a workspace with its own instructions and uploaded files, on every plan including Free. On Business, Enterprise and Edu plans it also offers workspace agents: shared assistants that can connect to your team's tools and run on a schedule. You'll still come across custom GPTs, the older version of the idea, but they're on the way out. Personal accounts can no longer create new ones, and OpenAI plans to retire them on all plans on 11 December 2026 (some Enterprise workspaces have until 11 February 2027), helping creators move them to plugins.
• Gemini calls it a Gem, the same concept in Google's world.
• Claude offers Projects (a workspace with its own instructions and uploaded knowledge) and Skills (a reusable folder of instructions Claude loads when a task needs it).
• Copilot offers agents, built in Microsoft's Copilot Studio, sitting inside your Microsoft 365 world.

Three ingredients recur every time: a name and description, a block of standing instructions (the important part; treat it as a very good prompt), and optional reference material the assistant always has to hand. Get those right and the assistant behaves consistently; get them vague and it stays vague, because it's the same model underneath with the same limits.

Where they differ

They are the same idea, but not interchangeable. The real differences are about where your data lives, who can use the result, and how much the assistant can reach beyond a chat.

• Knowledge. ChatGPT Projects and Claude Projects let you upload documents the assistant reads every time, useful for a house style guide or a product FAQ. Gems can hold files too, but simpler set-ups lean more on the written instructions.
• Reach. A Copilot agent can be wired into your actual Microsoft 365 content and, in Copilot Studio, into other systems. A ChatGPT workspace agent can likewise connect to apps such as Google Drive, Slack and SharePoint. Most Projects and Gems stay inside the chat unless you deliberately connect them to something. Reach is power and risk in the same move. An assistant that can act is one you must trust more.
• Sharing. Some can be shared with colleagues or published: ChatGPT Projects on its work plans, and workspace agents are built for it. That is handy and a privacy trap: never build confidential detail into an assistant you'll share.
• Where it lives. This is usually the deciding factor. If your documents are in Microsoft 365, a Copilot agent works on them in place. If your team runs on Google, a Gem fits. If you want a reusable expert you upload reference files to, a ChatGPT Project or a Claude Project suits.

A decision table

Use this to pick a starting platform. None of these are absolute. Availability and limits shift with your plan and change often, but the pattern holds.

Worked example: the same assistant, two tools

Say Fernway's office manager, Maya, tidies rough meeting notes into the same clean format every week. That is a perfect custom assistant, a repeated task with fixed rules. Here is the standing instruction, written once and reused whichever platform holds it.

You are a meeting-notes tidier for a UK operations team. When I paste rough notes, always return three things in this order: (1) a three-sentence plain-English summary a manager could read at a glance; (2) a Decisions list; (3) an Actions table with columns Action, Owner, Due date. Mark any action with no named owner as UNASSIGNED and any missing date as NONE. Use UK English and a neutral tone. Never invent an owner, a date, or a decision that isn't in the notes. If something is unclear, add it to a short 'Needs checking' list at the end.

Paste that same instruction block into a Gemini Gem's instructions field, and you have the same assistant in Google's world. Now compare them on one real input: the Fernway meeting notes, which deliberately contain unassigned actions and half-finished items.

Tidy these notes using your standing format. [paste the Fernway meeting notes]

You'll likely find both produce a sensible summary and a usable actions table. The differences are in feel: one may format the table more cleanly, one may be stricter about marking items UNASSIGNED, one may over-helpfully guess a due date you told it not to invent, which is exactly the behaviour your guardrail exists to catch. That last point matters: the instruction is doing the heavy lifting, not the badge on the box. When you find a platform that respects your guardrails and lives where your work lives, that's your answer.

Project instructions: You tidy rough meeting notes into our standard format (summary, decisions, actions table with Owner and Due date). Follow the uploaded house-style guide for tone and terminology. Never invent owners, dates or decisions; list anything unclear under 'Needs checking'.

Try it now

Common mistakes

• Learning it four times. Treating a Project, a Gem, a workspace agent and a Copilot agent as unrelated skills. They're one idea; the instruction-writing craft transfers wholesale.
• Starting something new on a custom GPT. If you can still create one, it's tempting, but OpenAI is retiring them. Build new assistants in a Project, or a workspace agent if your organisation uses them, so you don't have to rebuild later.
• Building one for a one-off. A custom assistant only pays back on genuine routines. For a single task, a normal chat is faster.
• Woolly instructions. "Be helpful" tells it nothing. The assistant is exactly as specific as the standing prompt you gave it.
• Trusting it more because it's "custom". This is the over-trust trap. A saved set-up feels official, but it's the same model with the same tendency to hallucinate. Guardrails in the instructions reduce bad behaviour; they don't remove the need to check facts and figures before you use them.
• Baking confidential data into a shareable assistant. If it can be shared, assume it will be. Keep sensitive detail out, and only put confidential company data into an employer-approved business tool.

Check your judgement

Two platforms can reuse the same instruction. Does that mean their sharing and permissions match?

Keeping current

Names, builders and plan limits move constantly across all four tools, and ChatGPT's in particular have changed this year. Check each vendor's own guidance rather than trusting a figure you read once: OpenAI's Projects in ChatGPT, ChatGPT workspace agents and GPTs in ChatGPT (which carries the retirement notice), Google's Gems in Gemini Apps, Anthropic's What are Projects? and Agent Skills overview, and Microsoft's Copilot Studio. Accurate as of 26 September 2026.