The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines an interaction surface via the ph.agent.request(prompt) API, which allows canvas viewers to pass natural language prompts to the agent responsible for authoring the canvas code.
Ingestion points: viewer-supplied text passed via the prompt argument in ph.agent.request (described in SKILL.md).
Boundary markers: The host platform requires the viewer to explicitly approve the prompt before execution and stages resulting changes as drafts for the creator's review (SKILL.md).
Capability inventory: The agent has the ability to retrieve source code via canvas-source-retrieve and stage new versions using canvas-draft-create (SKILL.md).
Sanitization: The host platform implements mandatory user confirmation and a review-and-promote lifecycle for all agent-generated modifications.
[COMMAND_EXECUTION]: The skill instructs the agent to use shell commands (specifically POST requests) to upload media files to dynamically generated URLs during the canvas asset management process.
[DATA_EXFILTRATION]: The skill facilitates the transmission of local image files to external URLs provided by the platform. This is a primary functionality for media management and includes explicit warnings to the agent against uploading sensitive data, credentials, or personally identifiable information.

building-canvases