Custom instructions vs Projects vs skills: which to use when

The three, side by side

From OpenAI's and Anthropic's help and pricing pages, read 5 November 2026.

Why not put everything in one place

It's tempting to pour every rule into custom instructions. The cost shows up quietly: your listing rules show up in a birthday-message chat, and the AI has to weigh 40 lines that don't apply to work out the 3 that do. Anthropic's advice on context for AI is to give the model the smallest set of high-signal information that gets the job done. Putting each rule where it applies is how you do that without thinking about it.

Projects have the opposite limit: what's in one project is invisible in the next. If you copy the same proposal rules into eight client projects, you're back to eight copies. That's the job a skill does: one method, used in every project.

A worked example

Illustration: a made-up business. Sam runs a two-person bookkeeping practice with 20 small-business clients. Her custom instructions had grown to a page. She sorts it:

• Custom instructions (5 lines): "I'm a bookkeeper in Ontario. Use Canadian spelling. Amounts in CAD. Ask before assuming a tax year. Short answers unless I ask for detail."
• A project per client (20): the client's chart of accounts, last year's notes, and their quirks ("owner pays the van insurance personally").
• Two skills: "Month-end email" (what to ask each client for, in her friendly tone, with a checklist) and "Expense question" (how to ask a client about an unclear transaction without sounding accusing).

Now in any client's project she types "month-end email for October" and gets the client's specifics from the project and her method from the skill. When she changed her document deadline from the 5th to the 7th, she edited one skill, not 20 projects.

Decide in three questions

1. Would this line be right in every chat you'll ever have? Yes: custom instructions. Keep that list to five or six lines.
2. Is it true only for one client, product or piece of work, and does it need files? Yes: that project's instructions, next to its files.
3. Is it how you do one kind of task, wherever it comes up? Yes: a skill, with a "use when" sentence so the AI picks it.
4. None of these: it's probably a one-off. Leave it in the chat.

Already pasting the same prompt every day? How to stop pasting the same prompt walks through moving it in 15 minutes.

What changes when a team is involved

For one person, the three places are a matter of tidiness. For a team, they decide whether everyone works the same way: custom instructions can't be shared at all, projects share on some plans only, and skills shared inside one AI product don't reach colleagues on the other one. Teaching a new hire your team's AI workflow shows how that plays out on someone's first day, and getting your team to use AI the same way covers sharing one version.

The AI skill below writes the instruction text itself: tell it what the assistant is for, and your AI drafts the role, the rules, the sources it may use, the format and the edge cases, short and in order of importance, ready to paste into a project.

Questions

What's the difference between ChatGPT custom instructions and Projects?

Custom instructions apply to all your chats. A project's instructions and files apply only to chats inside that project, and inside it they take priority over your custom instructions.

Claude Projects vs skills: which should I use?

A project is a place: one client or one piece of work, with its files and history. A skill is a method: how to do one kind of task, used in any chat or project where that task comes up. Many people use both: a project per client, and skills for the tasks they do for every client.

Does ChatGPT have skills?

Yes. On Business, Enterprise, Healthcare and Edu, skills work in chats, with admin controls (OpenAI's Skills help article). On Free, Go, Plus and Pro, OpenAI's pricing page (read 5 November 2026) lists Skills as a beta, and OpenAI's docs place standalone skills in the ChatGPT desktop app and Codex, and skills bundled in plugins in chats.