Skills

核心 · Key Idea

In one line: A Skill is a capability bundle — it packages everything one task needs (system prompt + tool list + examples + resources) into a single unit. Install a Skill and the agent "knows how" to do the task; uninstall it and the ability disappears.

What it is#

A few intuitive examples:

Image Skill — a prompt teaching the model how to write Stable Diffusion prompts plus a gen_image tool.
Weather Skill — a get_weather tool plus a "how to interpret weather" prompt.
PDF-reader Skill — a read_pdf tool plus examples on how to summarise papers.

Install N Skills into an agent and it gains all N specialisations simultaneously.

Analogy#

打个比方 · Analogy

The agent is the phone; Skills are apps.

Install the food-delivery app and it can order food.
Uninstall it and that ability vanishes.
Anthropic's Skills, Coze's "plugins", Dify's "Tool Sets" are all variants of the same idea.

Key concepts#

Skill ManifestManifest

A YAML / JSON: name + description + trigger conditions + resource list.

Bound ToolsBound tools

The function calls a Skill ships with (APIs, MCP servers, local code).

Skill PromptSkill prompt

Skill-private system addendum, spliced in only when the Skill is active.

ActivationActivation

Triggered by user choice / a manager-agent router / keyword match.

How it works#

One agent + many Skills = lightweight multi-agent — same LLM, costumes swapped per task.

Practical notes#

One Skill, one job. Don't cram "summarise documents" + "generate charts" into a single Skill. Single responsibility = better routing, easier debugging.
Skill descriptions drive routing. Like Function descriptions, write clearly "when to use, what it can do, what it can't."
Skills can compose. Complex capabilities ("write a product launch email") = writing Skill + chart Skill + proofreading Skill.
Hot-pluggable. A good Skill system adds / removes Skills at runtime without restarting the agent.
Audit / bill per Skill. Track Tokens and call counts per Skill so issues can be localised quickly.

Easy confusions#

Skill

**Capability bundle**: prompt + N tools + examples + resources.

Tool / Function

**A single atomic action**: one function.
Skill is the higher-level wrapper around tools.

Skill

**Same agent, costume change.**
Shared context, cheap, lightweight.

Multi-Agent

**Multiple independent agents** conversing.
Strong isolation and specialisation, but costly.