The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the developer to install the @cognite/app-sdk package via standard package managers (pnpm, npm, or yarn). This is a vendor-provided library necessary for the integration functionality described in the documentation.
[COMMAND_EXECUTION]: The skill utilizes standard development tools to read project configuration (package.json) and source files (App.tsx). These actions are intended to understand the project context before applying integration code and are consistent with the skill's primary purpose.
[INDIRECT_PROMPT_INJECTION]: The skill describes an integration pattern where an agent can ingest application data through 'resources' and execute tasks through 'actions'. While this creates an inherent surface for indirect prompt injection, the instructions include safety best practices, such as advising developers to include warnings for destructive actions and requiring user approval.
Ingestion points: Application state data is exposed via createAgentResource and contextual messages are sent to the agent via sendAgentMessage in SKILL.md.
Boundary markers: The instructions do not define specific delimiters for injected variables, but they encourage the use of structured JSON for data exchange.
Capability inventory: The skill facilitates the registration of application-specific tools (actions) that the agent can invoke, including potentially mutating operations.
Sanitization: The skill templates rely on the developer's implementation within the action handlers to validate or sanitize inputs provided by the agent.

integrate-fusion-agent