Skip to content

Skills ​

Skills give Rho reusable instructions for a task. Each skill lives in a SKILL.md file. Rho shows the model each skill's name and description through the skill tool. The model loads the full instructions only when it uses that skill, which leaves more context for your work.

In the interactive TUI, each load appears as a collapsed skill(<name>) card. Click the card or press Ctrl+O to read what the model received: the source and the instructions after the front matter. If a skill fails to load, the card shows the reason.

SKILL.md format ​

Create a directory for the skill and add a SKILL.md file. Start the file with YAML front matter between --- lines, then write the instructions in Markdown:

text
---
name: inspect-logs
description: Find and summarize errors in application logs. Use when asked to inspect or triage log files.
---

Read the relevant log files with the file tools, group errors by message,
and report the most frequent failures first with file and line references.

The file has three parts:

  • name (required): Use 1–64 lowercase letters, numbers, or single hyphens. The name must match the directory name. For example, inspect-logs/SKILL.md uses name: inspect-logs.
  • description (required): Tell the model what the skill does and when to use it. A specific description helps the model choose the right skill.
  • Body: Write the instructions after the front matter. Rho loads this text as written when the model uses the skill.

Front matter is real YAML, parsed by a maintained YAML library. Optional Agent Skills fields are supported and validated:

  • license: a license name or a reference to a bundled license file.
  • compatibility: up to 500 characters describing environment requirements.
  • metadata: a map of string keys to string values for extra properties.
  • allowed-tools: experimental, a space-separated list of pre-approved tools.

Rho also reads disable-model-invocation (a boolean). It is a client extension outside the Agent Skills field set. When true, the skill stays out of the prompt metadata and only loads through direct user invocation.

Where Rho looks for skills ​

Rho checks these locations in order. First match wins. Built-ins cannot be replaced by a user skill.

text
built-in skills (shipped with Rho)
~/.rho/skills/<name>/SKILL.md
~/.agents/skills/<name>/SKILL.md
<project>/.agents/skills/<name>/SKILL.md   (nearest directory first, up to the repository root)
<project>/.agents/plugins/<plugin>/skills/<name>/SKILL.md   (nearest directory first)
~/.agents/plugins/<plugin>/skills/<name>/SKILL.md

If Rho finds the same skill name more than once, it uses the first copy and logs the selected and ignored sources. You can't replace a built-in skill with a user skill. Skills in your home directory take priority over project skills. Loose skills take priority over plugin skills. Within a project, Rho starts at the working directory and searches up to the repository root. Plugin packages are described in Agent Plugins.

Add a skill ​

Loose skills do not need an install command. Choose where the skill should live:

  • ~/.rho/skills makes it available only to Rho.
  • ~/.agents/skills shares it with Rho and other agents that use this layout.
  • <project>/.agents/skills shares it with everyone who works in that repository.

Create a directory whose name matches the skill name, then add SKILL.md:

sh
mkdir -p ~/.agents/skills/inspect-logs
touch ~/.agents/skills/inspect-logs/SKILL.md

Open the new file and follow the format above. To add a third-party skill, copy its directory into one of these locations. Read its instructions before you use it.

Plugin-owned skills arrive through Agent Plugin packages (rho plugins install / link). Those skills keep the package as their owner and sit below every loose skill location in precedence. Loose skills and plugin-owned skills stay distinct in inventory output. Skills from a project's plugin packages activate only in trusted workspaces (RHO_TRUST_PROJECT_PLUGINS=1).

Load a skill yourself ​

You can load skills before the model responds instead of waiting for it to choose them:

  • Mention a skill as a whitespace-delimited $<name> token anywhere in a prompt, such as review this with $inspect-logs and $rho-config. Every known skill you mention loads once, in mention order, as one batch of skill calls before the first model request. Other text stays prose, so $20 and $HOME are left alone. In the interactive TUI, typing $ opens a skill palette, and tab completes the selected name.
  • Start a prompt with /skill:<name>. That skill loads first, followed by any skills the rest of the prompt mentions.

Mentions load only for prompts you send, including queued follow-ups. A steering message sent during a turn and the prompts that /init and the agent creator write leave $ text as prose. ACP hosts follow the same rule; see Agent Client Protocol.

Built-in skills ​

Rho includes three built-in skills:

SkillUse
rho-configConfigure Rho, including models, providers, credentials, aliases, permission mode, and direct config-file edits
rho-agent-creatorPower /agents create and /create-agent, including guided setup for claude-cli, cursor, and antigravity delegated runtimes
rho-workflow-authoringWrite and operate deterministic, resumable Starlark workflows

The agent creator requires direct user invocation and does not appear in the model's automatic skill catalog. The rho tool documents its diagnostics actions directly in its input schema.

See Tools and workspace for details about the skill tool. See Agents and delegation to learn how to define agents, including the agent definition schema.