It’s the perennial problem with AI coding agents: fiendishly brilliant with logic and language, but often too generic to grasp the nuances of real-world workflows.
This is what Anthropic is striving to address with Agent Skills, a new feature that lets developers define specialized capabilities, so Claude can follow established workflows and apply domain-specific expertise on demand.
While Claude already had certain “skills,” insofar as it was capable of generating things like spreadsheets or slide decks, with the new Agent Skills in tow, this capability is being opened up for developers to build their own reusable workflows.
It’s worth noting that Agent Skills aren’t unique to Claude Code, Anthropic’s AI coding agent. They work across the broader Claude ecosystem — from chat and document workflows in the main Claude app to automation pipelines via the API. In practice, a Skill could handle anything from drafting marketing copy to running a data-cleaning script.
But for this exercise, the focus is on Claude Code — because for developers, the prospect of a coding agent that can actually remember project context and reuse established workflows is where things get interesting. And yes, AI-infused software development is what we’re all about here at AI Native Dev.
To get started with Agent Skills, you’ll need an active Claude Pro, Max, Team, or Enterprise subscription. The feature can be enabled under Settings → Features → Agent Skills. Once active, Skills can be created through the built-in Skill Creator inside Claude’s chat or programmatically via the /v1/skills API.
Each Skill is a folder containing instructions, scripts, and resources, and every one starts with a SKILL.md file. This file defines, in plain-English, what the Skill does, how it should be used, and what tools or permissions it can access.
Developers can also restrict tool access, specifying whether Claude may run code, read files, or use certain APIs. It’s a form of scoped control that keeps Skills safe and predictable, ensuring they only operate within defined boundaries.
Inside Claude Code, Skills operate more like lightweight plug-ins than persistent memories. As you work, Claude scans the available Skills and selects whichever ones are relevant to the task at hand. According to Anthropic, it will load only the minimal information and files needed, which helps keep operations fast while still drawing on the required expertise.
Claude Code organizes Skills into three main scopes. Project Skills live inside a repository and become automatically available when that project is opened in Claude Code, making it easy to tie specific workflows or guidance directly to a codebase. Personal Skills are linked to an individual user’s account, carrying their own preferences and reusable workflows across projects and sessions. And Plugin Skills are bundled with Claude Code plugins — they’re automatically available when a plugin is installed and behave the same way as personal and project Skills, extending Claude’s capabilities through the plugin system without any extra setup.
Skills can also be combined when tasks overlap, reused across Claude apps, Claude Code, and the API, and even include executable code for steps where running a script is more reliable than text generation. Anthropic describes them as “custom onboarding materials” — essentially a way to package a team’s expertise so that Claude can act like a specialist, not a generalist.
In Claude Code, this plays out in tangible ways. A Testing Skill, for instance, might run unit tests via a defined script, flag any failures, suggest fixes, and even open a pull request if connected to GitHub. Another Skill could handle documentation updates or cleanup tasks, executing commands inside Claude’s sandboxed environment to keep a project tidy.
Following Anthropic’s announcement, developers were quick to weigh in on Agent Skills, some calling it a major step toward reusable AI workflows, others questioning how much it really changes day-to-day development.
Developer Simon Willison called Agent Skills “conceptually extremely simple,” noting that each is “a Markdown file telling the model how to do something… optionally accompanied by scripts that the model can run.”
Specifically, Willison highlighted token efficiency, an issue that is never far from mind in the AI software development community, where every extra token can mean slower responses and higher compute costs. At the start of a session, Claude scans the YAML frontmatter of each skill and reads only a short summary. “Each skill only takes up a few dozen extra tokens, with the full details only loaded in should the user request a task that the skill can help solve,” Willison said.
Elsewhere, Kaushik Gopal, principal engineer at Instacart, admitted his “first reaction was: ‘So what?’” — noting that tools like AGENTS.md, slash commands, nested instructions, and MCPs already attempt to define reusable agent workflows. What changed his mind, he said, was the way Anthropic approached the problem: “I don’t think the innovation is what Skills does or achieves, but rather how it does it that’s super interesting.”
Gopal frames this as part of Anthropic’s ongoing “push on context engineering as the next frontier.”
He points to progressive disclosure — the process by which Claude scans each Skill’s YAML frontmatter at startup, builds a lightweight index “with minimal token cost,” and loads full details only when relevant — as the real breakthrough. That dynamic context-loading model, Gopal argues, is “very token efficient… that’s the interesting development here,” and one that outperforms monolithic alternatives like AGENTS.md, which “load all context upfront” and risk bloating the context window at scale.
For all the debate, Agent Skills mark another quiet step in Anthropic’s bid to make Claude less of a clever prompt-follower and more of a programmable coworker. Whether developers see it as revolution or refinement, the direction is clear: context is the new code.
