Skills¶

Skills are one of Arc's main extension systems.

They let you add either:

executable tools written in Python
plain-text instruction files that shape agent behavior

What Skills Are For¶

Use skills when you want Arc to gain:

new local tools
reusable team instructions
domain-specific behavior
custom workflows without editing the core app

Built-In Skills¶

Arc already loads built-in skills at startup.

Examples include:

filesystem access
terminal execution
task delegation
browsing
browser control
code intelligence
scheduler tools
workflow tools
MCP gateway tools

You can inspect loaded skills from chat with:

/skills

Two Kinds Of User Skills¶

1. Python Skills¶

Put a .py file in ~/.arc/skills/.

Use this for custom tools with executable logic.

Example skeleton:

from arc.skills.base import Skill
from arc.core.types import SkillManifest, ToolResult, ToolSpec

class MySkill(Skill):
    def manifest(self) -> SkillManifest:
        return SkillManifest(
            name="my_skill",
            version="1.0.0",
            description="Does something useful",
            tools=(
                ToolSpec(
                    name="my_tool",
                    description="Run my custom action",
                    parameters={"type": "object", "properties": {}, "required": []},
                ),
            ),
        )

    async def execute_tool(self, tool_name, arguments):
        return ToolResult(tool_call_id="", success=True, output="Done")

Restart Arc and the skill is discovered automatically.

2. Soft Skills¶

Put a .md file in ~/.arc/skills/.

Use this for instruction-only additions such as:

writing rules
company conventions
domain language
persistent reminders about how the agent should behave

Example:

Always cite the source and date for factual claims.
Use concise bullet summaries before long explanations.
Prefer internal naming conventions used by the team.

Where Skills Apply¶

Bundled strategy text and user soft skills are added into the prompting layer used by Arc.

In practice, skills are a good fit when you want Arc to feel smarter everywhere without building a full external integration.

When To Use A Skill Instead Of MCP¶

Prefer a skill when:

the logic should live locally in Arc
you want a very small custom extension
you want direct Python control
you want instruction files, not just tools

If the tool already exists as an MCP server, MCP may be easier.