The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell scripts, scripts/new-canvas and scripts/tq, which use system commands like cp, mkdir, and open -a to manage diagram files and curl to communicate with the local tldraw server.
[DYNAMIC_EXECUTION]: The skill constructs and sends JavaScript code snippets to the tldraw application's local API (specifically the /api/search and /api/doc/:id/exec endpoints) to programmatically manipulate the canvas, create shapes, and manage pages.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from user prompts (infrastructure descriptions) and from the tldraw API (existing canvas shapes). It lacks explicit boundary markers or sanitization logic, creating a surface where malicious instructions embedded in the infrastructure description or canvas metadata could potentially influence agent behavior during the diagramming workflow.
Ingestion points: User-provided prose/lists in SKILL.md (Workflow Step 1) and canvas state retrieved via api.getShapes in references/shared/bridge.md.
Boundary markers: None identified in the instruction set.
Capability inventory: Execution of shell scripts (scripts/new-canvas, scripts/tq) and local network requests via curl.
Sanitization: No explicit sanitization or validation of input data is described beyond normalization to a JSON schema.
[EXTERNAL_DOWNLOADS]: The skill's documentation references the official tldraw website (offline.tldraw.com) for application setup and tldraw.dev for API documentation, which are well-known services associated with the skill's functionality.

deployment-diagram