pollen-robotics/reachy-mini-skills

Reachy Mini Skills

Curated skills for the Reachy Mini robot: small markdown documents that teach the robot's conversational LLM how to do something, or give it reliable knowledge about a topic - starting with itself.

A skill is context, not code: when loaded, its body is injected into the robot's system prompt for the current conversation. Think of them as the robot's bookshelf.

Layout

skills/
  <skill-name>.md    one file per skill

Skill format

Each skill is one markdown file with TOML front matter delimited by +++ (the same convention as conversation-app personality profiles):

+++
name = "my-skill"
description = "What this skill teaches, in one sentence. Use when [the concrete situations that should trigger it]."
+++

# My skill

The markdown body: the instructions or facts the robot loads into its
context when the skill is activated.

Front matter fields:

Field	Required	Notes
name	yes	kebab-case, specific (not helper/misc), must match the filename (without .md)
description	yes	see "The description is the discovery surface" below

The description is the discovery surface

When the robot browses skills, the description is the only thing the model reads to decide whether to load one. The body stays out of context until the skill is loaded (progressive disclosure), so a great body with a vague description is an unfindable skill.

Format - two sentences, third person, under 1024 chars:

1. What: the capability or knowledge the skill provides.
2. When: Use when ... with the concrete situations, keywords or questions that should trigger it.

# Good
description = "Factual reference on Reachy Mini's hardware: variants, motors, camera, microphones. Use when asked about the robot's body, sensors, price, or the differences between the two versions."

# Bad - gives the model no way to know when to load it
description = "Information about the robot."

Writing the body

Two kinds of skills, two shapes:

• Knowledge skills (facts about a topic): plain factual sections. Open with one line of guidance ("Use these facts when asked about...") and tell the robot to admit uncertainty rather than guess beyond the skill.
• Behavior skills (how to run an activity: a game, a routine, a ceremony): give numbered steps, the decision points ("if the player gives up, then..."), and one short example exchange - behaviors without concrete steps and examples drift.

In both cases:

• Small. The body lands in a system prompt where every token competes with the conversation. Aim well under ~600 words; split big topics into several skills. Assume the model is already smart: only write what it cannot know or must not improvise.
• Factual and sourced. Only state things you can back up. When a fact is uncertain, either drop it or say it is approximate. Avoid facts that expire quickly (versions, counts that change weekly); prefer stable truths or approximations ("about 80 emotions").
• Self-contained. No links the LLM would need to follow; the body must stand on its own.
• One term per concept. Pick a word ("app", not app/Space/program interchangeably) and keep it throughout.
• A default, not a menu. If several approaches exist, give the one to use and at most one escape hatch - not a list of options.
• English only, like all Reachy Mini metadata - the robot converses in any language on top of it.

Contributing

Open a pull request on this dataset adding one file under skills/. Review is the moderation gate: a skill's content goes straight into robots' system prompts, so PRs are read closely before merging.

License

Apache 2.0.