Skills

Skills are declarative descriptions of workflows, patterns, or domain knowledge that extend agent capabilities. They sit above tools in the abstraction hierarchy — a tool is a single action, a skill is a multi-step workflow.

Listing Skills

arc skills list
arc skills info code-review

Loading Skills

Load skills from a directory containing JSON skill files:

arc skills load ~/.arc/skills/
arc skills load ./project-skills/

Loaded skills are registered globally and available to all profiles.

Skill Format

Skill files are plain JSON with a name, description, and instructions field:

{
  "name": "code-review",
  "description": "Perform a thorough code review with security and performance checks",
  "trigger": "review",
  "instructions": [
    "Check for security vulnerabilities (injection, auth bypass, secrets in code)",
    "Check for performance issues (N+1 queries, unnecessary re-renders, memory leaks)",
    "Check for correctness (edge cases, error handling, type safety)",
    "Suggest improvements with code examples"
  ],
  "steps": [
    { "action": "read", "target": "changed files" },
    { "action": "analyze", "criteria": ["security", "performance", "correctness"] },
    { "action": "report", "format": "inline comments + summary" }
  ]
}

MCP-to-Skill Adapters

Any MCP tool can be wrapped as a skill using mcpToSkill(). This lets you compose MCP tools into higher-level workflows:

import { mcpToSkill } from '@axiom-labs/arc-core';

const skill = mcpToSkill({
  name: 'deploy-check',
  tool: 'run-tests',
  server: 'ci-server',
  preCondition: 'all tests pass',
  postAction: 'notify team',
});

Skillify self-improving

Skillify is a meta-skill where the agent creates new skill definitions from its own actions. When the agent performs a multi-step workflow, skillify can capture it as a reusable skill.

The generateSkillFromPattern() function detects repeated action sequences and proposes skill definitions:

# Skillify observes repeated patterns and suggests:
# "You've performed this 3-step deploy workflow 5 times.
#  Save as a skill called 'safe-deploy'?"

Stuck Detector

The stuck detector monitors agent actions for signs of looping or unproductive behavior:

• Repeated patterns — sliding window detection of repeated action sequences
• Jaccard similarity — compares recent actions to detect cycling
• Recovery strategies — cycles through different recovery approaches (rephrase, simplify, escalate)

const detector = new StuckDetector({
  windowSize: 10,
  similarityThreshold: 0.8,
});

// Returns recovery suggestion if stuck
const recovery = detector.check(recentActions);

When an agent appears stuck, ARC can automatically suggest a different approach or escalate to the user.

TUI Skills View

The TUI Skills view displays real skill data from the SkillRegistry.

Key	Action
r	Reload all skills from disk
i	Show detailed info for the selected skill

The view lists each skill's name, description, trigger keyword, and step count.

Web Dashboard Skill Management

The Web Dashboard Skills view provides skill management through a browser interface:

• Browse — view all loaded skills with trigger info and step details
• Reload — reload skills from disk (equivalent to arc skills load)
• Remove — unregister a skill from the registry

Changes are broadcast over WebSocket so connected clients see updates immediately.

TIP

The CLI provides the most complete skill management (load from arbitrary directories, register new skills). The TUI and Web Dashboard support reading the registry plus reload and remove operations.

Skill Registry

The SkillRegistry manages skill lifecycle:

• register — add a skill definition
• unregister — remove a skill
• findByTrigger — match skills by trigger keyword

Skills are loaded at startup from ~/.arc/skills/ and any directories specified in the profile or workspace config.