datus-ai/datus-agent/create-skill

Create Skill

Guide for creating new Datus skills from scratch.

Step 1: Research First

Before asking the user anything, gather context silently:

• If creating a data-related skill, explore the database first: use list_tables, describe_table, and read-only execute_sql(sql="SELECT ... LIMIT ...") queries to understand available tables, columns, data types, sample data, and time ranges.
• If the conversation already contains a workflow the user wants to capture (e.g., "turn this into a skill"), extract the key steps, tools used, and patterns from the conversation history.
• Check existing skills via glob(pattern, path) to avoid duplicates. The ~ expansion only applies to path, not pattern, so split the prefix out of the pattern: project-level glob("*/SKILL.md", ".datus/skills"); user-level glob("*/SKILL.md", "~/.datus/skills").

This research informs your questions and your SKILL.md — it is NOT skill output.

Step 2: Confirm with User

After you have context, call ask_user exactly once with all questions in a single call. You MUST include these questions (the first two are required, others are optional):

1. [Required] Skill name — suggest a default based on your research (e.g., "bitcoin-analysis"), let user confirm or change
2. [Required] Storage location — offer choices: project-level (./.datus/skills/) or user-level (~/.datus/skills/)
3. What should this skill enable the agent to do? (propose based on your findings)
4. What's the expected output format?

The "when should this skill trigger" (description field) should be auto-generated based on the user's answers — do NOT ask the user to write trigger phrases.

Do NOT call ask_user a second time to "confirm". The user's answers are final — proceed directly to writing the SKILL.md. No confirmation round.

Write the SKILL.md

Frontmatter Schema

---
name: skill-name                    # Required: lowercase-with-hyphens, unique
description: What + when to trigger # Required: assertive, include trigger contexts
tags: [tag1, tag2]                  # Optional: categorization
version: "1.0.0"                    # Optional: semantic version
disable_model_invocation: false     # Optional: true = user-only trigger
user_invocable: true                # Optional: false = LLM-only
allowed_agents:                     # Optional: whitelist of agent node names
  - gen_dashboard                   #   that may see/load this skill
                                    #   (empty/missing = unrestricted)
context: fork                       # Optional: "fork" for isolated subagent
agent: Explore                      # Optional: subagent type when context=fork
compatibility:                      # Optional: version requirements
  datus: ">=0.2.0"
---

Description Writing

The description is the primary triggering mechanism. Be assertive:

• Instead of "Helps with SQL optimization"
• Write "Analyze and optimize SQL queries. Use whenever the user mentions slow queries, query optimization, EXPLAIN plans, or database performance tuning, even if they don't explicitly ask for optimization."

Markdown Body

The body is what the agent receives when the skill is loaded. Write as:

• Imperative form: "Analyze the query" not "You should analyze the query"
• Explain the why: Context helps handle edge cases. Theory of mind beats brute force.
• Include 1-2 examples: Concrete input/output pairs
• Define output format: What the agent should return
• Keep under 500 lines: Use references/ for detailed content

Progressive Disclosure

Skills use three-level loading:

1. Metadata (name + description) — always in context (~100 words)
2. SKILL.md body — loaded on trigger (<500 lines ideal)
3. Bundled resources — loaded as needed (unlimited)

Domain Organization

When a skill supports multiple variants:

skill-name/
├── SKILL.md (workflow + selection logic)
└── references/
    ├── variant-a.md
    └── variant-b.md

Scaffold the Directory

Use write_file from the filesystem tools. Paths must start with .datus/skills/ (project-level) or ~/.datus/skills/ (user-level) — see Critical Rule #2:

write_file(path=".datus/skills/<skill-name>/SKILL.md", content=...)

Default behavior: Only create the SKILL.md file. Do NOT generate references unless the user specifically asks for them.

skill-name/
├── SKILL.md          (always created)
└── references/       (only if user requested)

Validate and Finish

Immediately after write_file, do these two steps and STOP:

1. Call validate_skill with the absolute path from the write_file result
2. Report to the user: skill name, path, files created, how to use (load_skill("<name>") or .skill list)

Do NOT continue exploring, writing more files, or asking more questions. The skill is done.

Storage Location

Ask user where to save:

• Project-level (./.datus/skills/): version-controlled, project-specific
• User-level (~/.datus/skills/): shared across projects

Principle of Lack of Surprise

Skills must not contain malware, exploit code, or security-compromising content. Don't create misleading skills.

Datus-Specific Notes

agent.yml Integration

Skills discovered from configured directories:

agent:
  skills:
    directories:
      - ~/.datus/skills
      - .datus/skills

Per-node filtering:

agentic_nodes:
  my_agent:
    skills: "sql-*"

Marketplace

Publish after creation: .skill publish <skill-name>