Paying Anthropic customers can now teach their Claude new tricks, which the company calls Skills.
Despite much talk about superintelligence, AI models may be clueless when it comes to interacting with specific applications. Sure, they can parse PDF text, but they probably don’t know how to fill out a PDF form.
Skills, or agent skills for jargon maximalists, offer a way to instill specific knowledge in Claude that he may not have in his distillation of training data. Anthropic has already deployed a few within Claude to handle common tasks like creating spreadsheets or presentations.
Now Claude’s paying customers (not the freeloader tier) can create skills tailored to their specific needs. Just be sure to subtract the work it took to create them when calculating the time saved by the AI.
A Skill consists of a directory with a SKILL.md file – a mix of YAML and Markdown – and optionally other resources such as text, scripts and data. This riot of instructions and executable code can be stored locally (~/.claude/skills/) or uploaded to the cloud for use with the Claude API.
Claude’s popup, when loaded, adds the metadata of the skills available in the system prompt. Subsequently, when Claude is asked to perform a relevant task, he will launch the appropriate skill by invoking the Bash tool to read the SKILL.md file. Once completed, the model should have the data needed to perform the desired task, such as interacting with content in a third-party app like Box or creating a PowerPoint presentation.
The AI model will do this through a process Anthropic calls progressive disclosure.
“Progressive disclosure is the fundamental design principle that makes agent skills flexible and scalable,” the company explains in its engineering blog on the subject. “Like a well-organized textbook that starts with a table of contents, then specific chapters, and finally a detailed appendix, the skills allow Claude to load information only when necessary.”
The result is that Claude does not unnecessarily process tokens for skills that will not be used, thereby reducing operating costs.
The skills also provide a way to return to executing programmatic code when a large language model would not be the right tool for the job. As an example, Anthropic cites the inefficiency of sorting a list via token generation: a coded sorting algorithm will accomplish the task much faster and more cheaply, and produce the same result every time.
Creating skills can be complicated for those who want to manually create the relevant files, but they are less cumbersome with Claude’s help. Claude integrates a “skills creator” which allows the creation of new skills thanks to the interactive jokes of the chatbot. There is also a Claude Skills recipe book for those who want to understand the creative process.
Speaking of malware, Anthropic warns that Skills poses some risks (much like giving Claude access to Bash). Malicious skills, the company warns, can introduce vulnerabilities or enable data exfiltration.
“We recommend installing skills only from trusted sources,” the company says. “When installing a skill from a less trusted source, audit it thoroughly before use. Start by reading the contents of the files bundled in the skill to understand what it does, paying particular attention to code dependencies and bundled resources such as images or scripts. Likewise, watch out for instructions or skill code that instruct Claude to connect to potentially untrusted external network sources.”
With this in mind, note that Anthropic says it hopes to allow AI agents to create their own skills. ®
