Both assistants added “skills.” Here is what that actually means for a six-person shop.

The word means something specific now

A skill is a folder of written instructions — sometimes with examples or a little code alongside — that an assistant picks up only when the job in front of it matches. Not a personality, not a subscription tier, not a chatbot with a different name. A written procedure that gets loaded when it is relevant.

What makes that more than a saved prompt is what happens when the skill is not being used. Most of the time only its name and a one-line description sit in front of the model; the instructions themselves get read when that description matches what you asked for. Which is why you can have twenty of these and pay nothing for the nineteen that have no bearing on today.

Anthropic published its version of the format as an open standard at the end of 2025, and a number of other tools have adopted it since. The practical upshot for an owner is that a procedure written down once is not necessarily welded to one vendor — though whether that holds for the specific product you pay for is a question for that product’s documentation rather than for us.

This is not “getting better at prompting”

A good prompt is typed by one person, once, and leaves the building when they do. A skill is the procedure written down: which system to look in first, what to check before drafting, what the finished thing has to contain, and what it must never do. The next person who asks gets the same sequence. So does the same person at four o’clock on a Friday.

For a business with two or three people doing the same recurring job, that consistency is the whole benefit. The assistant does not become cleverer. It stops inventing a slightly different process every time — which is the thing that makes AI output feel unreliable even when each individual answer was fine.

Anthropic publishes a ready-made set of these aimed squarely at ordinary small-business work: the Monday cash-and-pipeline brief, chasing overdue invoices, getting the month closed, working out which enquiries to call back first. They are a starting point rather than a finished system — they assume a business that keeps its books and its enquiries somewhere specific, and fitting them to how yours actually runs is most of the job.

The plan you are on decides whether you can see this at all

This is the part that quietly wastes people’s afternoons, so it is worth being blunt. At the time of writing, ChatGPT offers skills on its Business, Enterprise, Healthcare and Edu plans — not on Free, not on Plus, and not on Pro. An owner paying for Plus who reads an article about this and goes hunting for the menu will not find it, and there is nothing wrong with their account.

Claude goes the other way: skills are available across its plans including the free one, but they depend on a capability — code execution — that has to be switched on in settings, and that an administrator can switch off for everyone on a team plan. If the menu is visible but greyed out, that is almost always the reason.

Both companies move these lines, and neither asks us first. Check their own pages, linked at the foot of this one, before buying a plan on the strength of a feature you read about somewhere — this page included.

What actually goes wrong

Three things, and none of them are exotic. The first is that the skill never fires. The description is what the assistant matches against, so one described as “invoice helper” gets skipped by someone who asked it to “chase the overdue accounts.” A description has to say what the thing does and when to use it, in the words the person would really use — which is a writing problem, not a technical one, and it is where most of these quietly fail.

The second is a misunderstanding about data. A skill is instructions. It does not have your books or your customer list simply by existing — it gets those when you connect the underlying accounts. Connecting a live accounting or customer system to an AI assistant is a decision that deserves a written rule about who may do it and what may pass through it, not a checkbox clicked on a Tuesday. That policy is one of the things an internal agent build hands you, and there is a free one-page version in our AI use policy builder if you would rather write it yourself.

The third is trusting somebody else’s skill. A downloaded one is another person’s instructions, and possibly their code, running as you. OpenAI scans uploaded skills and will flag some for review or block them outright, which is a useful backstop and not a substitute for reading what you are installing. Treat one the way you would treat a spreadsheet macro emailed to you by a stranger.

So does a six-person shop need them?

Sometimes. The test we use has three parts: is the task recurring, could you write down how it is done, and can the judgement in it be separated from the typing around it? Chasing thirty overdue invoices passes all three — the deciding is yours, the drafting is not, and the sequence is the same every month. Quoting a job unlike any you have quoted before fails all three, and no amount of setup fixes that.

If one or two of your tasks pass, this is an afternoon of work rather than a project, and it is worth the afternoon. If none of them do, the honest answer is that a skill will not manufacture one and the time is better spent elsewhere. Which is the same conversation a first call with us is mostly made of: finding the two or three places this genuinely pays, and naming the places it does not.

Read it yourself

We would rather you check this than take our word for it, particularly the part about which plans include what. These are the primary sources, published by the companies themselves:

• Using skills in ClaudeAnthropic’s own instructions, including the capability that has to be switched on first.
• Skills in ChatGPTOpenAI’s help article — the authority on which plans have this and what an administrator controls.
• Agent Skills documentationThe technical format itself, for whoever ends up writing one.
• The small-business setThe ready-made routines described above, listed by the company that publishes them.

MAE-AI Solutions is an independent company that builds and runs AI for small businesses. We are not affiliated with, endorsed by, or a reseller for Anthropic or OpenAI. “Claude,” “Anthropic,” “ChatGPT” and “OpenAI” are trademarks of their respective owners and are used here only to identify the tools we work with. We build on other vendors’ AI tooling too, and we will tell you when something other than either of these is the better fit. Product capabilities, plan tiers and terms are set by the vendor and change — the links above are authoritative, this page is not.