My Custom Skill

My Custom Skill

This is a template for creating a custom agent skill.

How Skills Work

There are two valid skill patterns in the current architecture:

Pattern A: Prompt-Only (Recommended)

No script field in frontmatter. This SKILL.md is injected into the LLM's context when activate_skill is called. The LLM then uses built-in atomic tools (bash, codegen, workspace_read/write, glob, grep) to accomplish the task.

This is the primary extension mechanism — like Claude Code's slash commands.

Pattern B: Subprocess Script

Add script: scripts/my-script.py to frontmatter. The script is executed as a subprocess with --args-json and must communicate via stdout (JSON) and file I/O in the agent workspace (AGENT_WORK_DIR). Scripts cannot access the LLM or environment router — use this only for deterministic computation.

Behavioral Guidelines (Edit for Your Skill)

When this skill is activated:

1. Use codegen to query the environment for relevant information.
2. Use workspace_read / workspace_write to persist state.
3. Use bash to run any computation or data processing.
4. Call done with a summary when finished.

Example: A "Daily Journal" Skill

When activated, the agent should:

1. codegen with instruction: "What happened recently? Summarize recent events."
2. workspace_read path journal.jsonl to load previous entries.
3. workspace_write path journal.jsonl to append today's entry.
4. done with summary of what was journaled.