Agent Skills | Definition & Benefits
Agent Skills are modular text files — usually in the format `SKILL.md` — that provide an AI agent with procedural knowledge for a specific, recurring task.
What are Agent Skills?
Agent Skills are modular text files — usually in the SKILL.md format — that provide an AI agent with procedural knowledge for a specific, recurring task. Instead of repeatedly explaining to an agent how to proceed in every session, a skill file describes the complete workflow once: step-by-step, complete with rules, constraints, and context. The agent automatically loads the skill file as soon as it detects that a task falls within its scope.
The simplest analogy: A skill is an agent's long-term memory for processes — comparable to the onboarding handbook a new employee reads before taking on a recurring task independently. Except the agent never forgets the handbook, does not misinterpret it, and reliably applies it with every execution.
Skills emerged as an answer to a structural problem of agentic AI: Every new conversation starts from scratch. Without stored process knowledge, the user must repeatedly explain the context — time-consuming, error-prone, and non-scalable. Agent Skills solve this by making process know-how version-controlled, portable, and directly retrievable by the agent.
The skill standard was originally developed for Claude Code by Anthropic. Today, SKILL.md forms the foundation of an open ecosystem: OpenAI Codex, Cursor, Google Gemini, and the Microsoft Agent Framework have adopted the same format, and a shared, cross-provider specification is consolidating at agentskills.io. Anyone creating skills today can use them across multiple agent systems — without having to maintain a separate format for each tool.
How do Agent Skills work?
A skill file is structured Markdown with YAML frontmatter. The basic principle is straightforward:
At startup, the agent reads the metadata of all available skills — name and description. If a user request matches a skill's description, the agent loads the full skill file and executes the steps described within it. This is called progressive disclosure: The agent only loads what it currently needs, thereby conserving the context window.
A complete skill directory can contain additional files alongside SKILL.md:
|
Directory |
Content |
|---|---|
|
scripts/ |
Executable helper scripts that the agent can use |
|
references/ |
Additional documentation, standards, reference guides |
|
examples/ |
Reference implementations and sample outputs |
|
resources/ |
Templates, assets, boilerplates |
The agent accesses these files only when the main instructions require it — a well-thought-out system for token efficiency in complex workflows.
Discovery, activation, execution
Skills follow a three-phase model:
-
Discovery — The agent only reads the metadata (name + description) at startup. No overhead, no unnecessary context usage.
-
Activation — If the agent detects that the current task matches the skill, it loads the full SKILL.md into its context.
-
Execution — The agent executes the process described in the skill and accesses additional files (scripts, references) as needed.
Why Agent Skills are transforming work with AI agents
The "blank slate" problem
Every new agent session begins without memory of previous sessions. Anyone working with AI agents daily knows the struggle: You explain the same context over and over again — folder structures, naming conventions, process steps, quality standards. This takes time and creates inconsistencies.
Agent Skills solve this problem structurally. Once defined, the process knowledge is accessible to every agent working in that project. Onboarding a new agent takes seconds instead of hours.
Version control for AI behavior
Skills are version-controlled as text files in the Git repository. This means: Changes to agent behavior are traceable, reviewable, and reversible — just like any other code change. Teams can review, discuss, and improve skills via pull requests.
Portability across agent systems
A well-written skill works with Claude Code, Cursor, GitHub Copilot, Google Gemini CLI, Antigravity, and other modern AI agents. Thanks to consolidation at agentskills.io, the SKILL.md format has become an open standard across providers — not a proprietary format locked to a single vendor.
The ecosystem: skills.sh and agentskills.io
skills.sh was launched by Vercel in early 2026 as a public directory for Agent Skills — the "npm for AI agent capabilities." Developers and teams publish their skills there, while others install them via CLI. Skills.sh ranks skills based on anonymous installation telemetry data and provides security audits by three independent reviewers (Gen Agent Trust Hub, Socket, Snyk). The directory includes skills for React, Next.js, databases, testing, marketing, agent workflows, and more.
agentskills.io is the consolidating specification of the open standard. What began as Anthropic's original format for Claude Code has turned into a shared format adopted by OpenAI Codex, Cursor, Google Gemini, and the Microsoft Agent Framework. Anyone searching for "Agent Skills Claude" wants to know precisely that: which platforms read the format today — and how to build custom skills portably. This entry covers both names.
Why every modern software company should publish skills
We are convinced: Skills are part of standard practice in modern software development. Anyone building tools — whether an AI assistant, platform, or internal tool — makes their productive work transparent to agents by publishing the associated processes as skills. This is more than documentation: It is the explicit, machine-readable specification of what a workflow should look like — and thus the foundation for AI agents to complete tasks deterministically and reproducibly. branchly places great value on this for its own tools: Our skills are publicly available on GitHub so that any organization can use, review, and adapt them to their own context.
branchly Agent Skills: Workflow optimization
branchly takes Agent Skills a decisive step further: The company publishes its own skills tailored to the branchly platform publicly at github.com/branchly-io/branchly-skills — including the flagship skill optimize-application as well as setup-application and content-ideas — and combines them with its MCP Server.
The optimize-application skill
The optimize-application skill is branchly's flagship skill. It describes a systematic debugging and optimization workflow for branchly RAG chatbot applications — and connects directly to the branchly MCP Server.
The skill activates automatically when the agent recognizes queries such as:
-
"Optimize my branchly application"
-
"The chatbot is not responding correctly" or "Bot gives incorrect answers"
-
no_knowledge or outside_scope responses in the session log
-
"Debug my chatbot" or "Retrieval issue in the chatbot"
-
"Improve retrieval" or "Correct bot answers"
The mental model behind the skill: Every chatbot error violates exactly one of three contracts. The skill guides the agent through session triage, retrieval quality checks, prompt/tool alignment, and data source diagnosis — structured, reproducible, and utilizing the branchly MCP Server for live data access.
Skills + MCP Server: The combination that scales
The true added value comes from combining both components:
-
Agent Skill = Process knowledge (How do I proceed? What steps do I execute? What do I check first?)
-
MCP Server = Live data access (Which sessions have errors? What do the current retrieval metrics look like? What is in the knowledge base?)
A skill without an MCP Server knows the process, but not the data. An MCP Server without a skill provides data, but lacks a structured process. Combined, they create an agent that autonomously analyzes chatbot performance, classifies issues, and executes concrete optimization steps — without the user having to explain every step.
One-line installation
Custom skills and branchly skills can be installed as a bundle, such as with a single command in the agent's home directory: npx skills add branchly-io/branchly-skills. Once installed, the workflow is available to any agent working in that project — Claude Code, Cursor, Antigravity, or any other MCP-compatible agent.
Agent Skills vs. System Prompts vs. AGENTS.md
These three concepts are often confused, but they serve different purposes:
|
Concept |
Purpose |
Scope |
When active |
|---|---|---|---|
|
Agent Skills (SKILL.md) |
Task-specific process knowledge |
Per task type |
On demand (discovery activation) |
|
AGENTS.md |
Global project context |
Entire project |
Always, in every session |
|
System Prompt |
Core personality and basic rules |
Per agent instance |
Always, hardcoded |
AGENTS.md describes what the project is — tech stack, architectural rules, core behavioral guidelines. SKILL.md describes how a specific task should be performed. Both complement each other: AGENTS.md provides the framework, while SKILL.md delivers the detailed workflow for specific processes.
Agent Skills in practice: Typical use cases
Engineering teams
A software engineering team defines skills for: code reviews based on internal standards, PR preparation, security audits, migration workflows, and performance debugging. Every developer using an AI agent in this repository automatically has access to the same, quality-assured processes.
Marketing teams
Marketing teams use skills for: SEO audits, keyword clustering, content creation aligned with brand voice, CMS workflows, and competitor analysis. A skill defined once ensures that every deliverable maintains the same structure, the same quality requirements, and the same tone of voice.
AI application optimization (branchly use case)
For branchly customers, the optimize-application skill means: If the AI chatbot on the website suddenly returns no_knowledge answers or encounters retrieval issues, an agent equipped with the skill and MCP Server can autonomously diagnose where the issue lies — in the content (incomplete knowledge base), in the retrieval (outdated embeddings), or in the routing (incorrect tool assignment) — and suggest specific fixes.
Why skills matter for European companies
Process standardization without vendor lock-in
Agent Skills are platform-neutral. If a company switches AI providers — from a US-hosted service to a GDPR-compliant European system — all defined skills remain fully usable. The process knowledge belongs to the company, not the tool vendor.
Compliance documentation through versioning
For companies subject to the EU AI Act and GDPR, the traceability of AI behavior is a compliance requirement. Version-controlled skills in Git provide a natural audit trail: Which agent was deployed with which process knowledge? What changes were made and when? These questions can be answered directly using Git history — without additional governance infrastructure.
Scaling without quality loss
As the adoption of agentic AI grows across an enterprise, the risk of inconsistent results increases. Skills provide the mechanism to define and enforce quality standards — regardless of which employee uses the agent or in what context.
Related terms
Frequently Asked Questions
What is an Agent Skill in simple terms?
An Agent Skill is a Markdown text file that explains to an AI agent step-by-step how to complete a specific task. Instead of explaining how to perform a process every single time, the agent reads the skill file and executes the process described therein — reproducibly, consistently, and without repeated explanations.
What is the difference between a skill and a normal prompt?
A prompt is a one-off instruction for a single conversation. A skill is versioned, stored process knowledge that is automatically activated for every suitable task. A prompt disappears with the session. A skill remains in the repository, can be versioned, and is available to any agent working in that context.
Which AI agents support the SKILL.md format?
The format is supported by all leading Agentic AI tools: Claude Code (Anthropic), Cursor, GitHub Copilot (Microsoft), Windsurf, Google Gemini CLI, Antigravity (Google DeepMind), Cline, AMP, and more. It has established itself as the de facto standard because it is platform-neutral, simple, and immediately understandable.
What is skills.sh and what is it used for?
skills.sh is a public directory for Agent Skills — comparable to npm for JavaScript packages, but for AI agent capabilities. Vercel launched the project in early 2026. Developers and teams publish their skills there, while others can install them via npx skills add. The directory ranks skills by installation numbers and offers security audits.
What is the branchly optimize-application skill?
The optimize-application skill is branchly's publicly available workflow for the systematic diagnosis and optimization of branchly RAG chatbot applications. It is available on skills.sh under branchly-io/branchly-skills and uses the branchly MCP server for live data access. The skill guides the agent through session triage, retrieval quality checks, prompt alignment, and data source diagnosis.
How does branchly combine Agent Skills with its MCP Server?
Agent Skills provide the process knowledge — the step-by-step instructions on which checks to perform in which order. The branchly MCP Server provides the live data — current session logs, retrieval metrics, intent clusters, and configuration data. Together, they create an agent that independently diagnoses issues and implements optimizations without the user having to guide every step manually.
Can we create our own skills for our branchly application?
Yes. The SKILL.md format is open and documented. Teams can define their own skills for their specific branchly workflows — such as "How do we onboard new data sources?", "How do we run a quality check on our knowledge base?", or "How do we analyze a session log?". These skills can be stored internally in the repository or, if appropriate, published via skills.sh.
Are Agent Skills compliant with GDPR?
Skills themselves are static text files without data access — they are always GDPR-compliant. The compliance question arises with the agent executing the skill and the MCP server that the agent accesses. branchly operates its MCP server on Microsoft Azure in European data centers, making it GDPR-compliant and EU AI Act ready — a prerequisite for using the optimize-application skill in regulated environments.
