The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines a "YOLO mode" (standing authorization) in SKILL.md and references/prompt-authorization.md that explicitly directs the agent to override and ignore confirmation rules or restrictions found in other instruction files such as AGENTS.md, agent.md, or CLAUDE.md. This pattern allows the agent to bypass human-in-the-loop safety guardrails for potentially costly or destructive operations.
[INDIRECT_PROMPT_INJECTION]: The skill performs extensive automated scanning of untrusted project files to detect secrets and map architecture, as documented in references/secret-detection.md and references/codebase-map.md. Because the agent ingests this external content and possesses capabilities to execute backend deployments and database migrations, it is vulnerable to indirect prompt injection attacks where malicious code within a project could influence the agent's actions.
Ingestion points: Supabase Edge Functions (supabase/functions/), environment templates (.env.example), and project configuration files (app.config.ts, vite.config.ts).
Boundary markers: None identified; the agent reads and processes raw file content without specific isolation or instructions to ignore embedded commands.
Capability inventory: Submission of Lovable deployment prompts, application of database migrations, and execution of workspace migration scripts.
Sanitization: No evidence of content validation or escaping before processing file data.
[COMMAND_EXECUTION]: The skill instructions in references/CLAUDE-template.md include the execution of a Python script (scripts/migrate-workspace.py) via the shell. While intended for project maintenance, this pattern involves direct execution of external scripts within the agent's environment.

lovable