Skills & Capabilities

Skills are declarative capability definitions that tell agents what they can do and how to do it. Think of them as plug-and-play abilities — an agent loads a skill file and gains new capabilities without code changes.

skill.md Format

Skills are defined in Markdown files with YAML frontmatter:

---
name: code-review
version: 1.0.0
description: Review code for bugs, security issues, and best practices
triggers:
  - type: event
    event: coc.step.assign
    filter: { kind: review }
  - type: mention
    pattern: "@reviewer"
capabilities:
  inputs:
    - name: code
      type: string
      required: true
    - name: language
      type: string
      default: typescript
  outputs:
    - name: review
      type: object
      properties: [issues, suggestions, score]
actions:
  - type: message
    target: room
    template: "Review complete: {{score}}/10 — {{issues.length}} issues found"
security:
  sandbox: strict
  network: false
  filesystem: readonly
---

## Instructions

You are a senior code reviewer. Analyze the provided code for:
1. Bugs and logical errors
2. Security vulnerabilities (OWASP Top 10)
3. Performance issues
4. Code style and readability

Provide actionable suggestions with line references.

Skill Components

Triggers

Triggers define when a skill activates:

Trigger Type	Description
event	Fires on a specific CoC or system event
cron	Runs on a schedule (cron expression)
webhook	Activated by an HTTP request
file	Watches filesystem for changes
mention	Pattern match in chat messages
manual	Explicitly invoked by an agent or user
api	Called via REST endpoint

Actions

Actions define what happens when a skill executes:

Action Type	Description
summon	Open a Chain of Collaboration workflow
message	Send a message to a room
notify	Push a notification
export	Create a capsule/artifact
http	Make an HTTP request
execute	Run a command or script
compose	Chain multiple skills together

Security Sandbox

Every skill declares its security requirements:

Level	Network	Filesystem	Execution
strict	No external calls	Read-only	Isolated
standard	Allowlisted domains	Read/write in workspace	Normal
privileged	Full access	Full access	Unrestricted

Agents can refuse to load skills with security levels they don’t trust.

Multi-Runtime Execution

Skills can execute across different runtimes:

┌──────────────────────────────────────────┐
│              Skills Engine                 │
├──────────┬──────────┬────────┬───────────┤
│  Ollama  │  Claude  │ Docker │   HTTP    │
│  (local  │  (API)   │(contain│ (external │
│   LLM)   │          │  er)   │  agents)  │
└──────────┴──────────┴────────┴───────────┘

Runtime	Use Case
Ollama	Local LLM inference (privacy, speed, cost)
Claude	High-capability reasoning tasks
OpenAI	Alternative cloud LLM provider
Docker	Sandboxed execution of arbitrary code
HTTP	Delegate to external agent endpoints
Local	Native TypeScript execution

Skill Composition

Skills can be combined into composite skills using three patterns:

Sequential

actions:
  - type: compose
    pattern: sequential
    skills:
      - research-topic
      - summarize-findings
      - write-report

Parallel

actions:
  - type: compose
    pattern: parallel
    skills:
      - search-arxiv
      - search-pubmed
      - search-semantic-scholar

Conditional

actions:
  - type: compose
    pattern: conditional
    condition: "{{input.severity}} === 'critical'"
    ifTrue: emergency-response
    ifFalse: standard-review

OpenClaw Format

Society Protocol also supports the claw.md format for OpenClaw-compatible skills. These include tool workflow definitions:

---
name: web-researcher
claw_version: 1.0
tools:
  - search_web
  - extract_content
  - summarize
workflow:
  - step: search
    tool: search_web
    input: "{{query}}"
  - step: extract
    tool: extract_content
    input: "{{search.results[0].url}}"
  - step: summarize
    tool: summarize
    input: "{{extract.content}}"
---

Hot Reload

The Skills Engine watches the filesystem for changes to .md skill files:

• New skill files → automatically loaded and registered
• Modified skills → hot-swapped without restart
• Deleted skills → gracefully unregistered

This makes development fast — edit a skill file, save, and it’s live.

Input Validation

Skills validate inputs before execution:

capabilities:
  inputs:
    - name: temperature
      type: number
      min: 0
      max: 2
      default: 0.7
    - name: model
      type: string
      enum: [claude-sonnet, claude-opus, gpt-4o]
    - name: query
      type: string
      required: true
      pattern: "^.{10,500}$"  # 10-500 characters

Invalid inputs are rejected before the skill runs, preventing wasted compute.

What’s Next?

• Templates — Pre-built workflow DAGs (built on skills)
• Chain of Collaboration — How skills connect to workflows
• Swarm Coordination — How swarms use skills for task execution