> ## Documentation Index
> Fetch the complete documentation index at: https://skillcliprotocol.io/llms.txt
> Use this file to discover all available pages before exploring further.

# How skills work

## The agent loop

When an agent receives a task, it follows this flow:

```
User message
    ↓
Agent scans available skills (reads SKILL.md descriptions)
    ↓
Agent selects matching skill(s)
    ↓
Agent reads full SKILL.md for usage instructions
    ↓
Agent runs commands / scripts as described
    ↓
Agent reads output and responds to user
```

There is no capability negotiation, no handshake, no initialization sequence. The agent reads a file and acts on it.

## Discovery

Skills are discovered through the filesystem. An agent knows about skills in its configured skill directories. When a task arrives, the agent scans skill descriptions to find relevant ones.

```
~/.skills/
  github/SKILL.md
  weather/SKILL.md
  database/SKILL.md
```

This is analogous to how a developer scans tool documentation — read the description, decide if it's relevant, read the details if so.

## Execution

When a skill includes executable scripts, the agent runs them as child processes:

```bash theme={null}
# Agent decides to use the weather skill
./weather/run.sh "Tokyo"
# stdout: Tokyo: ☀️ +15°C
```

The agent captures stdout and uses it as context for its response.

### Environment variables

Skills can receive configuration through environment variables. This keeps secrets out of command-line arguments (which appear in process listings).

```bash theme={null}
# Agent sets env before calling
GITHUB_TOKEN=ghp_xxx ./github/run.sh pr list
```

### Exit codes

* `0` — success, stdout contains the result
* Non-zero — failure, stderr contains the error message

## Composition

Skills can reference other skills. A "deploy" skill might call a "git" skill and a "docker" skill internally. This is just scripts calling scripts — no orchestration framework needed.

## What SCP does NOT define

SCP intentionally avoids specifying:

* **Transport protocols** — there is no transport; skills are local files
* **Authentication handshakes** — use environment variables or config files
* **Capability negotiation** — the agent reads markdown; it understands what it reads
* **Lifecycle management** — no connections to open or close
* **Streaming protocols** — stdout is already a stream