Skills can be created in three ways:
| Mode | Source | Description |
|---|
| Skill Folder | Local directory or zip | Upload a pre-built skill folder containing SKILL.md |
| Prompt | prompt (+ optional json_schema) | Auto-generate SKILL.md and schema.json from a text prompt |
| Session | session_id | Auto-generate SKILL.md from a chat session’s history |
From Skill Folder
Upload a local skill folder directly. The folder must contain a SKILL.md file — the skill name and description are parsed from its YAML frontmatter automatically.
# Upload a local skill folder directly (zips and creates in one step)
vlmrun skills upload ./my-skill
# Override name/description from SKILL.md frontmatter
vlmrun skills upload ./my-skill --name "invoice-extraction" --description "Extract structured data from invoices"
from vlmrun.client import VLMRun
from pathlib import Path
client = VLMRun(api_key="<VLMRUN_API_KEY>")
# One-step: zip, upload, and create a skill from a local directory
skill = client.skills.create_from_directory(
directory=Path("./my-skill"),
)
print(f"Created skill: {skill.skill_name} (type={skill.type})")
# Override name/description from SKILL.md frontmatter
skill = client.skills.create_from_directory(
directory=Path("./my-skill"),
name="invoice-extraction",
description="Extract structured data from invoices",
)
create_from_directory handles zipping the folder, uploading the archive via the Files API, and creating the skill in one call. It returns an AgentSkill with type="skill_reference" that you can pass directly to any endpoint that accepts skills.
The folder should follow the skill directory structure:
my-skill/
├── SKILL.md
├── schema.json
├── vlmrun.yaml
└── resources/ (optional)
Use the skill folder method when you need full control over the skill’s instructions, schema, and execution configuration. Use the prompt method for quick prototyping.
From Prompt
Generate a skill automatically from a text description and optional JSON schema:
# From a text prompt
vlmrun skills create --prompt "Extract invoice_id, date, and total_amount from invoices."
# With a JSON schema file
vlmrun skills create --prompt "Extract invoice data" --schema schema.json
from vlmrun.client import VLMRun
client = VLMRun(api_key="<VLMRUN_API_KEY>")
skill = client.skills.create(
prompt="Extract invoice_id, date, and total_amount from invoices.",
json_schema={
"type": "object",
"properties": {
"invoice_id": {"type": "string"},
"invoice_date": {"type": "string", "format": "date"},
"total_amount": {"type": "number"}
},
"required": ["invoice_id", "invoice_date", "total_amount"]
}
)
print(f"Created skill: {skill.id} ({skill.name})")
SKILL.md with instructions derived from your prompt and a schema.json from the provided JSON schema.
From Chat Session
Generate a skill from an existing chat session’s conversation history:
vlmrun skills create --session-id "<session-id>"
from vlmrun.client import VLMRun
client = VLMRun(api_key="<VLMRUN_API_KEY>")
skill = client.skills.create(session_id="<session-id>")
print(f"Created skill: {skill.id} ({skill.name})")
