The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell scripts (scripts/new-canvas, scripts/tq) to perform file operations and interact with the tldraw desktop application. These scripts use standard system tools like mkdir, cp, open, and curl to manage the canvas lifecycle.
[DYNAMIC_EXECUTION]: The skill generates JavaScript code at runtime to be executed within the tldraw application context via the /api/doc/:id/exec endpoint. This is used to programmatically create and style shapes, arrows, and frames based on user-provided diagram descriptions.
[INDIRECT_PROMPT_INJECTION]: The skill processes natural language descriptions or transition lists to generate state diagrams, creating a surface for indirect prompt injection.
Ingestion points: User-provided prose or transition lists processed in the Workflow section of SKILL.md.
Boundary markers: The instructions mandate a normalization step where the agent must convert input into a strict JSON schema ({ states, transitions }) before proceeding.
Capability inventory: The skill has the ability to execute shell commands locally and run arbitrary JavaScript within the tldraw application via a local API.
Sanitization: The requirement to parse input into a specific data model serves as a validation layer, reducing the risk of the agent directly obeying instructions embedded in the user data.
[CREDENTIALS_UNSAFE]: The scripts/tq script reads a server.json file from the application's configuration directory (e.g., ~/Library/Application Support/tldraw/) to retrieve a per-launch bearer token. While this is a sensitive credential for the local API, it is managed dynamically and is not hardcoded in the skill source code.

state-diagram