Claude Agent Skills, One Month In: Portable, Not Yet a Standard

A month into Claude Agent Skills, the scene repeating in team channels is small: someone drops a skill folder in a channel, and because it is just files you unzip it and let your agent load it. No install, no account, no plugin review. That frictionlessness is the point of the format, and the exact moment to slow down. One month after Anthropic introduced Agent Skills on October 16, 2025, the honest status is narrower than the hype and still genuinely useful: a skill is portable across Claude apps, Claude Code, and the API, it is not yet a cross-agent open standard, and the low friction that makes it easy to share is not the same as low responsibility to review.

Most coverage this fall kept answering the wrong question, redefining what a skill is for the tenth time. A month in, the questions that actually decide anything are two: how far does a skill really travel, and what are you running when you load one.

What Anthropic actually shipped in October

The launch post describes Agent Skills as folders of instructions, scripts, and resources that Claude loads when they are relevant. Three properties are worth holding onto. Skills are composable, so more than one can apply to a task. They are efficient, because the agent loads only what a job needs instead of everything at once. And they can include executable code for the parts where a deterministic script is more reliable than asking a model to produce the answer token by token.

That last property is the quiet one, and it reframes everything downstream. A skill is not only prose for the model to read. It can carry scripts the agent runs. A folder that behaves the same way on a repeatable operation does so because a script handled it, not because the model happened to get it right that time.

There was a signal before the announcement, too. Anthropic’s public skills repository was created on September 22, 2025, about three weeks ahead of launch. A reference collection that exists before the product tells you the format was meant to be read, copied, and learned from, not sealed inside a UI.

Portable across Claude, not across everything

Here is the precise claim, and precision is the whole point of a dated snapshot. At launch, skills work across Claude apps, Claude Code, and the API. That is portability across one company’s surfaces, and it is real. Write a workflow once, and the same folder applies whether you are in a Claude app, in Claude Code, or calling the API.

What has not been announced is a cross-platform, cross-agent open standard. The temptation is to round the story up to universal portability, because the format is just files and files go anywhere. Resist it. “It is only a folder” describes the packaging, not the runtime. As of this writing the credible statement is the narrower one, and it is still strong: a real packaging pattern that travels across the Claude family, with a public reference repo to learn from.

“Just files” is exactly why review is your job

Now the part the frictionlessness hides. The launch post is explicit that the API form of Agent Skills requires the Code Execution Tool beta, because skills need a secure environment to run in. Sit with that. It is a direct statement that skills are executable, not passive text, and that running them safely takes an actual sandbox. The same guidance tells you to use skills only from trusted sources, precisely because a skill can give Claude the ability to execute code.

So the folder format lowers distribution friction to nearly zero. It does not lower review responsibility at all. Those are two different dials, and treating a shared skill like a saved prompt because both arrive as text is how a portable format becomes a portable problem. A saved prompt cannot reach for your shell. A skill with a script can.

The good half of the trade is that plain files are reviewable in a way a hosted plugin never was. You can open the folder, read the instructions, and read the scripts to see what the agent will be told to do before it runs. Concretely, one month in, that review is a short habit: read the SKILL.md to see the workflow, open anything under scripts/ to see what actually executes, and check the parts that reach the network or the shell. The format hands you the ability to do that. It does not do it for you, and nobody else does it either.

What the format has, and what it is still missing

Take stock. The market has a credible packaging pattern, executable where it needs to be, portable across the Claude family, with a public reference repository that predates the launch. What it does not yet have is a cross-platform open standard, a shared way to say this exact skill runs on any compliant agent.

That gap is where curation earns its keep. Aaria keeps a deliberately small catalog for this reason: a portable format still needs evidence that a workflow actually works and clear boundaries on what it touches, and neither of those falls out of the file format for free. Portability answers how a skill moves. It does not answer whether the skill is any good, or safe, which is what an admission standard is for.

The right read for November 2025 is neither hype nor dismissal. The format is real, and it travels further than a prompt ever did. The open standard is not here yet. Build on the first fact, review every folder as if it can run code, because it can, and do not assume the second fact before it arrives.

FAQ

Are Claude Agent Skills a Claude-only feature or an open standard? As of November 2025 they are a Claude feature with real portability across Claude apps, Claude Code, and the API. A cross-platform open standard has not been announced, so treat cross-agent portability as an expectation, not a shipped fact.

What is inside a Claude Agent Skill? A skill is a folder of instructions, scripts, and resources. Anthropic describes it as composable and efficient, loading only what a task needs, and it can include executable code for work a deterministic script handles better than a model.

Do skills run code? Yes. The launch post says the API form requires the Code Execution Tool beta because skills need a secure environment to run, and it warns you to use skills only from trusted sources. A skill is executable, not passive text.

How should I review a skill before running it? Because the skill is plain files, read the SKILL.md for the workflow, open the scripts/ directory to see what executes, and check anything that reaches the network or the shell. Then decide whether you trust the source, the same way you would with any dependency.