Skills

Skills are procedural knowledge. They tell the model how to handle a situation, when to use tools, and what quality bar to apply.

They are different from tools. Tools do things. Skills teach the agent how to work.

MCP guide skills follow the same rule. When createHarness({ mcp }) includes a server guide, Little Harness stages a guide under .agents/skills/<server-id>-mcp/SKILL.md; that file explains when and how to use the server, but it does not grant access. MCP access comes from the configured gateway tools, usually mcp_list_tools and mcp_call_tool.

Skill Shape

A skill can be a directory with a SKILL.md file:

SKILL.md should contain frontmatter and instructions:

---
name: customer-briefs
description: Use when preparing renewal, expansion, or risk briefs for account teams.
---

Before writing the brief, inspect recent support incidents, product usage,
contract dates, and open account notes. Prefer evidence-backed bullets.

Using Skills

const harness = createHarness({
  host: localHost(),
  model,
  system: "Help account teams prepare high-signal customer work.",
  tools: { searchDocs, queryCustomers, searchIncidents },
  skills: [
    skill("./skills/customer-briefs"),
    skill("https://github.com/anthropics/skills", {
      skills: ["frontend-design"],
    }),
  ],
});

The runtime puts skill names and descriptions in the prompt, then stages full skill bodies under .agents/skills/<name>/ from the harness working directory for on-demand reading:

cat .agents/skills/frontend-design/SKILL.md

With runtime: { bash: false } the agent has no shell to read them with, so the harness adds a load_skill tool instead. It takes a skill's name and returns its SKILL.md, or another file of the skill when asked, and the skill list in the prompt says to open skills with it.

Harness memory adds a built-in memory skill. A skill of your own named memory replaces it.

The helper that registers a skill is documented in skill().

What Belongs In A Skill

Use skills for:

• domain vocabulary
• decision procedures
• report formats
• tool usage rules
• review checklists

Do not use skills for secrets, user-specific memory, or large raw datasets. Put persistent app data behind tools or Persistent Dirs, and what the agent learns in memory.