Claude Agent Skills — how a SKILL.md file replaces the custom agent
Anthropic introduced Agent Skills on October 16, 2025, and published the format as an open standard on December 18, 2025. The Claude Code 2.1.0 release on January 7, 2026 added hot-reloading — skills now activate without restarting the session. The promise is plainly stated: a Markdown file with frontmatter replaces the work that used to go into building a full custom agent.
What a Skill actually is
- A directory containing a
SKILL.mdfile. The frontmatter only requiresnameanddescription; the rest is Markdown. - Progressive disclosure: Claude initially reads only the description. The body loads only when a task actually matches.
- Portable: the same skill works in Claude.ai, Claude Code, the API, and in MCP-aware third-party agents.
- Bundleable: scripts, templates, and examples can sit in the same folder — Claude pulls them in as needed.
What used to be true
Until autumn 2025, anyone wanting to specialize Claude for a recurring task had three options, all of them effortful: build a custom agent with system prompt and tool bindings, write an MCP server, or copy the entire logic into every prompt. Anthropic itself describes the previous state as “fragmented, custom-designed agents for each use case” — one custom agent per use case, each with its own setup and maintenance.
In daily practice this meant: every new task (building PowerPoint slides, populating PDFs, evaluating Excel reports) came with its own mini-project. Sharing knowledge — across team members or projects — meant repos, READMEs, and onboarding sessions.
What’s true now
1. One file instead of a project. The simplest Skill example in Anthropic’s own documentation is a folder with a single file inside: SKILL.md. The frontmatter block contains a name and a description; the body is prose with instructions and examples for how Claude should apply the skill. It only grows from there if you want it to: extra scripts, templates, or reference markdown files just live in the same folder. Anthropic puts it this way: “Skills are a simple concept with a correspondingly simple format. This simplicity makes it easier for organizations, developers, and end users to build customized agents and give them new capabilities.”
2. Progressive disclosure saves tokens. Claude initially loads only the frontmatter descriptions of all available skills. It only reads the body once the current task matches. That keeps the context small and lets you have dozens of skills available in parallel without artificially bloating every request. That is the difference from “stuff everything into the system prompt” — skills are retrievable, not permanently loaded.
3. Open standard, not a proprietary island. With the December 18, 2025 release Anthropic published the format as an open standard. Other vendors and tools can support the same format. That makes skills a serious complement to MCP — Anthropic’s framing: skills extend MCP rather than replace it, covering the “workflow layer” while MCP stays the “tool layer”.
4. Hot reload since January 2026. With Claude Code 2.1.0 (January 7, 2026), skills in ~/.claude/skills or .claude/skills activate immediately, without restarting the session. That removes the last friction point that previously slowed down skill development.
Why this shifts the tooling landscape
Skills solve a concrete problem that costs serious time in practice: knowledge a team accumulates over months — “this is how we write pull request descriptions”, “this is how we build slides for quarterly reviews”, “this is how we run conversion-tracking audits” — otherwise stays trapped in heads, Notion pages, or scattered READMEs. A Skill is the first standard that captures that procedural knowledge in a form an agent can reliably apply.
For agencies like ours this matters because repetitive work can finally be cleanly abstracted away without writing custom tooling. A Skill called “SEO audit, our way” or “newsletter brief from a client call” is a Markdown file that lives in the repo, is versionable, and works anywhere a Skill-aware agent runs.
What you can do now
If you currently maintain custom agents or system prompt collections: check whether they can be expressed as Skills. Pre-built skills for Office documents (PowerPoint, Excel, Word, PDF) are already available across the Claude surface — they make solid templates for your own.
If you work in a team: set up a skills structure in a shared repo (.claude/skills/<area>/SKILL.md). That makes knowledge versioned, reviewable, and identical for everyone on the team — whether they use Claude Code, Claude.ai, or the API.
If you have MCP servers: keep them. Skills don’t replace MCP, they complement it. MCP stays the right tool for external integrations (databases, APIs, external services); skills are the right tool for the workflow knowledge that drives Claude.
Entdecke mehr
Claude Code: How to enable Fast Mode for Opus 4.6 — and what it costs
Fast Mode for Opus 4.6 is now extra-usage only. Here's how to enable it — and what it really costs you.
BlogWhy AI Models Find Different Code Problems
Three frontier models, the same 1000-line script, three different finding lists — and why that very spread makes multi-orchestration strong.
GlossarEnsemble / Multi-Model Orchestration
Ensemble means combining several deliberately varied LLM runs or models whose findings complement each other. Multi-model orchestration drives these runs via orchestrators with sub-agents, so the union of results is larger than any single run.