On 18 December 2025, Anthropic published a file format.
A file format is only an agreed way of writing something down so that different software can all read it. A .csv is one. So is a .pdf. Nobody owns it, and anybody can use it.
Within 48 hours Microsoft had put this one into VS Code and OpenAI had put it into both ChatGPT and Codex. By March this year, 32 tools from companies that compete with each other were all reading the same file from the same folder, including Google’s Gemini CLI, JetBrains’ Junie and Amazon’s Kiro.
That is the part worth sitting with. Rival companies do not normally adopt each other’s formats. They build their own, because a format you own is a customer who cannot easily leave. When the biggest names in the industry all pick up the same one inside three months, it is because they had the same problem, and none of them thought it was worth fighting over.
The format is called a skill. It sits between a prompt and an agent. Those three words get used as if they mean the same thing, and they are three different purchases.
What is a prompt?
A prompt is a request with context attached.
You type what you want and everything the model needs to know in order to get it right. The brief, the audience, the tone, the format, the constraints. Then you do it again tomorrow, from scratch. If you want the longer version of why that context matters more than the wording, I wrote about why context beats prompts every time.
A prompt starts from nothing every time. It has no memory of yesterday’s session, and no way to reach your files or your systems unless you paste them in yourself. That is not a flaw. It is what makes it fast.
Prompts are free, immediate and endlessly flexible. They are the right tool for anything exploratory, anything one-off, and anything you are still working out.
Their limit is not intelligence. It is repetition. If four people in your business need the same output, the context gets typed four different ways and you get four different results. If you need it monthly, you retype it monthly, and it drifts every time.
The fix is not a better prompt. It is to stop retyping.
What is an AI skill?
A skill is a reusable recipe that packages the context so nobody has to retype it.
In the open standard Anthropic published, a skill is a folder. Inside it sits a file called SKILL.md, holding a name, a description and the instructions. Alongside that you can bundle scripts for the calculations that have to be exact, reference documents, and templates.
Mike Bayly at The AI Corner describes it in the plainest terms I have seen. A reusable recipe that AI uses to complete a task or workflow. The instructions, your company context, code for exact calculations, and where to go to retrieve information.
The mechanism underneath is worth knowing, because it explains why skills scale where prompts do not. The agent loads only the name and description of every skill it has at startup, which costs almost nothing. When a task matches, it reads the full instructions. Only then does it pull in the scripts or documents it needs. So you can keep dozens of skills on hand without paying for them until they are used.
What changes commercially is that your standards stop living in someone’s head. The person supplies the task. The skill supplies the context. Same brief, same output, whoever asks.
What is an AI agent?
An agent decides its own next step.
That is Anthropic’s own distinction, and I wrote it up in full in what is an AI agent, actually. A workflow runs steps a person wrote in advance. An agent directs its own process and keeps control of how the job gets done. It thinks, calls a tool, reads what came back and decides again, until it has enough.
A skill makes the output consistent. An agent decides what the next task should be.
Are you already using an agent?
Almost certainly, and it is worth knowing that before you buy anything.
Agents are not exotic and they are not only something you purchase. They are already inside the tools you use. Claude Code sends out its own sub-agents to work on separate parts of a job at the same time. ChatGPT has an agent mode. If you have used either this week, you have used an agent.
That changes the question. It is not whether agents work, or whether they are worth having. They do and they are. It is which kind you are dealing with.
An agent built into a tool you already pay for is a feature. Somebody else carried the build cost and somebody else supervises it. You get the benefit for the price of a subscription, and you should use it.
An agent built for your business is a project. You carry the build, and you carry the checking for as long as it runs. Same word, different thing entirely, and it is the second one the rest of this post is about.
The same job, three ways
Take the monthly customer newsletter, because almost every business has one and almost every business dreads it.
As a prompt, you open a chat and paste in the month’s news, your tone notes, who the list is, the length you want and the two products you are never allowed to discount. You read the draft, fix it, send it. Next month you do all of that again.
As a skill, you type “write November’s newsletter” and hand over the month’s news. The recipe already holds the tone, the audience, the format, the never-discount rule and where the sales figures live. You still read the draft before it goes.
As an agent, it looks at what actually happened in the month, decides which of it is worth telling anyone, writes the thing and schedules the send. Notice which word did the work there. Decides. Nobody told it what November’s news was.
What actually changes as you move up?
Not capability. The same model sits underneath all three.
What changes is how much of the checking you have handed over.
With a prompt, you read every word before it goes anywhere. With a skill, you still read the output, but you are no longer checking whether the right context went in, because the recipe handles that. With an agent, you are no longer reading each step at all. You are reading a result that arrived by a route you did not choose.
That is the actual purchase. Each step up, you trade a bit of checking for a bit of speed.
And checking, it turns out, is the expensive part. McKinsey costed agentic workflows last week. For a customer service agent in banking, tokens came to 20% to 25% of the variable running cost. Human oversight came to 70% to 75%.
The model is the cheap bit. The watching is the bill.
What does each one cost?
Roughly, and in the shape that matters rather than in dollars.
A prompt costs you the minute it takes to type and the minute it takes to read the answer. Nothing else. There is no setup and nothing to maintain.
A skill costs an afternoon to write properly and very little after that. It is a text file. It does not need monitoring, it does not act on anything, and if it produces something wrong you see it before it goes anywhere, exactly as you would with a prompt.
An agent costs the build, then it costs supervision for as long as it runs. Somebody has to define what it may decide, watch what it decided, and catch the calls that were wrong. McKinsey found production agents often need adjusting every couple of days as models and systems change around them.
There is also a volume test, and for most New Zealand businesses it is the one that settles the argument. Agent economics improve with scale and only with scale. McKinsey costed an agent onboarding 2,500 customers a year at $10,000 to $15,000. Doubling the customers took it to $15,000 to $20,000, not $30,000, because the fixed cost stays still while the volume climbs. Run the same agent across 50 jobs a year and you are carrying that whole fixed cost on 50 transactions.
Which one do you actually need?
Start at the bottom and only move up when the one below stops working.
- Use a prompt when the task is one-off, exploratory, or still taking shape. Most work is this.
- Use a skill when you find yourself typing the same context more than twice, or when more than one person needs the same output to come out the same way. This is where most businesses have real gains sitting untouched, and it is the least talked about of the three, because it is a text file and nothing about that sounds impressive.
- Use an agent when decisions genuinely have to be made mid-task, the volume is high enough to spread the fixed cost, and you have someone whose job includes watching it. All three, not two.
If your version of the task happens a handful of times a month, it is not an agent problem yet.
The one-page version
All three side by side, what each one costs, and the four questions that settle it. One page, free, no email required.
Where should a New Zealand small business start?
With the thing you retype.
Open the last month of your AI use and find the context you have supplied more than once. Your tone, your customers, your pricing rules, the numbers that must never be estimated, the format your reports come in. That list is your first skill, and it will take an afternoon.
Do that before anyone sells you an agent. Not because agents are bad, but because a skill is what an agent would need underneath it anyway, and you will have built it for the cost of an afternoon rather than a quarter.
Gartner reckons around 130 of the thousands of agentic AI vendors are the real thing, and expects more than 40% of agentic AI projects to be cancelled by the end of 2027 on cost, unclear value and inadequate risk controls. The businesses least likely to be in that 40% are the ones who worked up from the bottom rather than jumping to the top.
Which part of your work are you still retyping every week? If you want to talk one through, let’s talk.
Sources
- Agent Skills, open standard originally developed by Anthropic, released 18 December 2025.
- Anthropic, Building Effective Agents, 19 December 2024.
- McKinsey QuantumBlack, Where AI agents pay off, 24 August 2026.
- Gartner, Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 25 June 2025.
- Mike Bayly, The AI Corner NZ, 16 August 2026.
