The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a development framework for creating canvases that ingest data from external providers (e.g., GitHub, MCP servers) and PostHog's own data warehouse.
Ingestion points: Data is pulled into the canvas context via ph.connectors.call, ph.loadInsight, and ph.query methods.
Boundary markers: The skill recommends progressive loading with state-specific skeletons and requires explicit user gestures (clicks) for triggering any side-effect actions (ph.actions).
Capability inventory: Canvases can perform authenticated writes to PostHog (ph.actions), store persistent state (ph.state), and capture analytics (ph.capture).
Sanitization: The instructions emphasize using "Proven Queries" (saved insights) and typed query nodes rather than raw HogQL to ensure data integrity and verifiability, which mitigates risks associated with processing untrusted data.
[SAFE]: The skill utilizes a platform-provided SDK (@posthog/canvas-sdk) that operates within a host-managed sandbox. Network access and tool execution are strictly limited by a capabilities configuration that prevents unauthorized resource access.
[SAFE]: The documentation includes explicit security best practices, such as warning against storing PII or secrets in team-visible state (ph.state) and restricting external navigation to PostHog domains via ph.openExternal.

querying-canvas-data