Agent Skills are reusable instructions you define once, so you don’t have to repeat them every time you interact with an LLM.
They are modular capabilities that extend Claude’s functionality. Each Skill packages instructions, metadata, and optional resources (scripts, templates) that Claude uses automatically when relevant.
Technically, it is a folder with a SKILL.md file containing instructions, optional scripts, and reference files.
Skill Types
- Pre-built Skills : Provided by Anthropic for common tasks like PowerPoint, Excel, Word, PDF creation
- Custom Skills : Created by individuals to package specific tasks as per there needs and workflows
Once you write the skills, they are available across Claude’s agent products:
- Claude.ai provides a web and mobile chat interface where you can easily use Anthropic’s pre-built Skills or upload your own custom Skills as ZIP files via the settings.
- Claude API is for developers and teams who want to integrate Claude directly into their own applications. Skills can be uploaded once to the API and shared across the entire workspace — every team member gets access automatically.
- Claude Code is a CLI for developers working inside their codebase. This is where Skills are most powerful — you create a Skill as a folder on your machine, run claude using CLI, and it is picked up automatically. No upload, no configuration.
- Claude Agent SDK is for teams building autonomous AI agents programmatically in TypeScript or Python. Skills plug in through filesystem-based configuration.
Across all four surfaces, Skills serve the same purpose — they give Claude reusable, domain-specific expertise so it knows exactly what to do without being told every time.
Architecture & Working
Let us understand the architecture and working by this custom skill , which will help to summarize any given git repo.
The Skill consists of two parts.
- First, a SKILL.md file that contains the instructions it tells Claude what to do when someone asks to summarize a repository.
- Second a script in Python or Java that do the actual work — connecting to GitHub, fetching the repository structure, README, dependencies, and source layout, then detecting the tech stack.
When you ask claude like “Summarize this repo”, Claude reads the Skill instructions, runs the scripts on your machine, collects all the repository data, and passes it to the LLM as context.
The LLM then reasons over everything and returns a clean, structured summary.

The key components of this architecture are
- Agent Configuration — Tells Claude what Skills exist and when to trigger them
- VM / Filesystem — Where the actual Skill files and scripts live and execute
- SKILL.md — The instructions Claude follows once triggered
- Scripts — Program which will have programs and instructions to execute
- LLM — Reads the instructions and script output, generates the final answer
Now let us look at exactly how Claude orchestrates all of this under the hood.
- Agent Configuration (Boot): Claude initializes by scanning the
SKILL.mdlibrary. It loads only the names and descriptions into its system prompt, creating a “map” of available capabilities without cluttering the context window. - User Request (Trigger): You provide a natural language prompt (e.g., “Summarize this repo”). Claude parses the intent and performs a semantic match against its internal map of skill descriptions.
- Instruction Loading (Just-in-Time): Once a match is found (e.g.,
github-repo-summariser), Claude fetches the full instructions from the local filesystem. This keeps the prompt lean until the specific skill is needed. - Script Execution (The “Doing”): Claude follows the
SKILL.mddirectives to run local scripts (e.g.,fetch_repo.py). These scripts interact with external tools like the GitHub CLI or Python runtimes. Only the resulting data is piped back to Claude. - Reasoning (The “Thinking”): Claude synthesizes the fresh data with the skill’s specific logic. The LLM processes the input to generate a structured, intelligent response.
Sharing & Distribution
How a Skill reaches your team depends on the platform:
- Claude Code: Most flexible. Keep Skills personal in
~/.claude/skills/or commit them to a project root.claude/skills/so anyone who clones the repo gets them automatically. - Claude API: Centralized. Once uploaded, the Skill is available to everyone in the workspace.
- Claude.ai: Manual. Every team member must upload their own copy; no central admin distribution exists yet.
Runtime & Network Access
- Claude Code (The Powerhouse): This environment offers full network and system access. It operates with the same permissions as any other program on your machine, allowing Skills to call external APIs, access the internet, and execute system tools like
gh(GitHub CLI) orjava. - Claude API (The Vault): This is an isolated environment with no network access. Skills cannot make external calls or install new packages at runtime; they are limited strictly to pre-installed libraries. This makes it highly secure but less flexible for data-fetching tasks.
- Claude.ai (The Variable): Network access here is plan-dependent. Depending on your administrative settings and subscription tier, access may be full, partial, or entirely restricted. It is best suited for simple, self-contained logic.
In the next part of the article we will see how to build skill for a git summarizer with scripts written in Java


