The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a local command-line interface binary (lovable) to manage project states, authorization, and tool execution. It invokes various subcommands including status, authorize-url, and call-tool.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core functionality of reading untrusted external data.
Ingestion points: The agent is instructed to read project source code, messages, and knowledge base content (code.read, projects.read, project_knowledge.read) from the Lovable platform as specified in manifest.yaml and SKILL.md.
Boundary markers: The instructions do not define clear boundaries or provide warnings to the agent to treat data from these ingestion points as potentially adversarial instructions.
Capability inventory: The skill possesses significant capabilities, including editing projects (projects.edit), running database queries (database.query), and publishing code changes (projects.publish).
Sanitization: There is no evidence of content sanitization or filtering before the ingested data is processed by the agent's logic.
[DYNAMIC_EXECUTION]: The skill employs a dynamic tool-calling mechanism (lovable call-tool). It fetches and executes tool definitions from a remote provider catalogue (lovable.dev) using the lovable call-tool command. While the instructions indicate that tools are reviewed before being made available, this architecture involves executing remotely defined logic at runtime.

lovable