Skip to Content
ConfigurationSkills

Skills

Skills are on-demand capability packages for agents. Each skill has a description (so the agent knows when to use it), instructions (injected when loaded), and optionally a set of tools that become available upon loading. This keeps agent context lean until a skill is actually needed.

Defining Skills

skill "web_scraping" {
  description  = "Load when you need to extract data from web pages"
  instructions = load("./skills/web_scraping.md")
  tools        = [plugins.playwright.browser_navigate, plugins.playwright.browser_snapshot]
}
 
skill "data_analysis" {
  description  = "Load when you need to analyze datasets"
  instructions = "Use the dataset tools to query, filter, and aggregate data..."
}

Attributes

AttributeTypeRequiredDescription
descriptionstringyesShort description — shown to the agent so it knows when to load the skill
instructionsstringyesDetailed instructions injected as a system prompt when loaded. Use load() for long content.
toolslistnoTools that become available to the agent when the skill is loaded

Assigning Skills to Agents

Agents reference global skills via the skills attribute:

agent "researcher" {
  model       = models.anthropic.claude_sonnet_4
  personality = "Thorough"
  role        = "Research specialist"
  tools       = [builtins.http.get]
  skills      = [skills.web_scraping, skills.data_analysis]
}

The agent’s system prompt will list available skills with their descriptions. At runtime, the agent calls load_skill when it decides a skill is needed.

Agent-Scoped Skills

Skills can be defined inside an agent block, making them available only to that agent:

agent "assistant" {
  model       = models.anthropic.claude_sonnet_4
  personality = "Helpful"
  role        = "General assistant"
 
  skill "summarizer" {
    description  = "Load when asked to summarize content"
    instructions = "Summarize content in exactly 3 bullet points. Be concise."
  }
 
  skill "translator" {
    description  = "Load when asked to translate text"
    instructions = load("./skills/translator.md")
    tools        = [plugins.translator.translate]
  }
}

Agent-scoped skills are implicitly available — they don’t need to be listed in skills = [...]. They can also be used on mission-scoped agents:

mission "research" {
  commander { model = models.anthropic.claude_sonnet_4 }
 
  agent "worker" {
    model       = models.anthropic.claude_sonnet_4
    personality = "Efficient"
    role        = "Worker"
    skills      = [skills.web_scraping]
 
    skill "mission_helper" {
      description  = "Load for mission-specific analysis"
      instructions = load("./skills/mission_analysis.md")
    }
  }
 
  agents = [agents.worker]
  task "analyze" { objective = "Analyze the target" }
}

How Skills Work at Runtime

  1. The agent sees available skills listed in its system prompt (name + description)
  2. When the agent decides a skill is needed, it calls load_skill with the skill name
  3. The skill’s instructions are injected as a system prompt (persistent, survives compaction and pruning)
  4. The skill’s tools (if any) are added to the agent’s available tools
  5. The agent receives a confirmation and proceeds using the skill’s instructions and tools

Rules

  • A scoped skill name must not conflict with any global skill name
  • Two agents can each define a scoped skill with the same name (independently scoped)
  • load_skill is available to agents only, not commanders
  • Skills persist through compaction and pruning because they are injected as system prompts
  • instructions can be an inline string or loaded from a file via load()
Last updated on
AgentsTools