Skip to content

Skills, Templates, and Capabilities

An agent's behaviour comes from three separate layers, and it helps to keep them straight:

  • Tools are what an agent can do — the operations it's allowed to call (search the web, read email, write an Odoo record). Tools are governed by permissions.
  • Instructions are who the agent is and the rules it follows — its SOUL.md and AGENTS.md. See Instructions vs. Memory.
  • Skills are how the agent uses its tools for a particular job — a reusable workflow guide for a capability, like "monitor a market" or "triage a mailbox."

A skill doesn't grant a tool and it isn't a personality. It's a playbook: a SKILL.md file that walks the agent through doing one thing well with the tools it already has.

Skills are attached through templates. When you create an agent from a template that includes a skill, Pinchy:

  1. Materializes the skill body into the agent's workspace at skills/<skill-id>/SKILL.md.
  2. Emits skills: ["<skill-id>"] in that agent's OpenClaw configuration.
  3. Lets OpenClaw load the SKILL.md into the agent's system prompt automatically — the workspace copy takes precedence over any bundled skill of the same name.

Because Pinchy writes the skill files at config-regenerate time — alongside the AGENTS.md and SOUL.md bootstrap files — the skill travels with the agent and stays in sync with its configuration.

| Skill | Attached by | What it teaches | | ------------ | --------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------- | | web-search | Market & News Monitor | A workflow for researching a topic with the web-search tools and reporting back. | | email | Email Assistant templates | A workflow for triaging, reading, and drafting mail with the email tools. |

OpenClaw ships dozens of bundled "desktop" skills (1Password, Apple Notes, and the like) that make no sense for a governed enterprise agent. Pinchy's skills: [...] entry is an allow-list: it enables exactly the Pinchy-authored skills an agent's template attached, and nothing else. That's the same fail-closed principle Pinchy applies to tools — list only what's trusted, and deny the rest by default.