Skills

A skill is a bundle of domain-specific instructions, with optional Python scripts and resource files: how to fill a PDF form, the house style for a report, the steps of a support procedure. The agent sees only the list of its skills, with their descriptions, and loads a skill’s full instructions when a task calls for it. Dozens of skills cost a few lines of prompt until they are used.

On disk

One directory per skill, holding a SKILL.md file:

my_skills/
├── weather/
│   ├── SKILL.md        # frontmatter + instructions
│   ├── scripts/        # optional: Python functions, registered as tools
│   │   └── tools.py
│   └── resources/      # optional: files the agent can read on demand
│       └── stations.csv
└── quarterly-report/
    └── SKILL.md

SKILL.md

---
name: weather
description: How to answer weather questions: which source to use, units, and the format of a forecast. Load it for any question about weather, temperature or forecasts.
license: MIT
allowed-tools: get_forecast
---

# Weather

1. Call `get_forecast` with the city in English.
2. Give temperatures in Celsius, then Fahrenheit in parentheses.
...

Key	Meaning
`name`	the skill’s name, what the agent loads it by
`description`	when to load it: the model chooses from the descriptions, so say what the skill is for and the words that call for it
`uri`, `license`, `compatibility`, `allowed-tools`	optional, informative
`metadata`	free key-value pairs

The body is the instructions, as long as they need to be.

Giving skills to an agent

from gwenflow import Agent, ChatOpenAI
from gwenflow.skills import SkillsDirectory

skills = SkillsDirectory("./my_skills")

agent = Agent(
    name="Assistant",
    llm=ChatOpenAI(model="gpt-5-mini"),
    skills=skills.skills,
)
print(agent.run("What's the weather like in Paris?").content)

SkillsDirectory loads every sub-directory that has a SKILL.md and skips the others. The agent then gets:

a compact list of the skills (name and description) in its system prompt;
three tools: list_skills, load_skill(skill_name) which returns the full instructions with the lists of resources and scripts, and read_skill_resource(skill_name, resource_name);
the tools defined by the skills’ scripts.

The skills block is not rendered when the agent has a fixed system_prompt.

Scripts

Each .py file in a skill’s scripts/ directory can provide tools in one of two ways:

a module-level tools list of Tool instances, used as is;
otherwise, every public function with a docstring and type annotations, wrapped with Tool(fn).

scripts/tools.py

def get_forecast(city: str, days: int = 3) -> str:
    """Return the weather forecast for a city.

    Args:
        city: Name of the city, in English.
        days: Number of days, 1 to 7.
    """
    ...

The tools are registered on the agent automatically, once each: passing tools=skills.get_tools() as well creates no duplicates. A script is imported once per skill, under its own module name.

Script files are imported and executed as Python. Only load skills from directories you trust.

Skills in code

A skill does not need a directory:

from gwenflow.skills import Skill

style = Skill(
    name="house-style",
    description="The writing rules for anything sent to customers. Load it before writing an email or a reply.",
    content="- Address the customer by name.\n- Short sentences.\n- Never promise a date.",
)

agent = Agent(name="Support", llm=llm, skills=[style])

Skill.from_directory(path), Skill.from_yaml(file) and Skill.from_dict(data) load one skill.

With sub-agents

An orchestrator with sub-agents can hand its skills down: each spawn_agent call names the skills the sub-agent may load.

Get started

Concepts

Models

Agents

Multi-agent

Extend

Memory

Flows

Knowledge

Production

Project

On disk

Giving skills to an agent

Scripts

Skills in code

With sub-agents

​On disk

​Giving skills to an agent

​Scripts

​Skills in code

​With sub-agents

On disk

Giving skills to an agent

Scripts

Skills in code

With sub-agents