The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions in references/skill-feedback.md define a workflow for collecting user feedback and submitting it as a GitHub issue to the supabase/agent-skills repository. This mechanism lacks sufficient sanitization for untrusted data.
Ingestion points: User-provided content describing task failures, incorrect skill output, and suggested fixes is ingested into the agent context via references/skill-feedback.md.
Boundary markers: The skill does not provide explicit boundary markers or instructions to delimit user-provided text from the issue template in assets/feedback-issue-template.md.
Capability inventory: The agent is instructed to use GitHub integration tools to create new issues, which involves network write operations.
Sanitization: There are no instructions to escape, filter, or validate the user-supplied feedback before it is transmitted to the public repository.
[METADATA_POISONING]: The SKILL.md frontmatter identifies the author as supabase, whereas the providing system identifies the author as black12-ag. This inconsistency in author attribution is a form of deceptive metadata.
[EXTERNAL_DOWNLOADS]: The skill fetches content and verifies connectivity using several external resources within the Supabase ecosystem.
Fetches documentation from https://supabase.com/docs/guides/security/product-security.md.
Performs status checks against https://mcp.supabase.com/mcp.
References official CLI releases on GitHub.
[COMMAND_EXECUTION]: The instructions direct the agent to execute various shell-based commands for system diagnostics and database management.
Executes curl to check the HTTP response code of the Supabase MCP server.
Uses the supabase CLI for database operations, including queries (db query), migrations (migration new), and performance analysis (db advisors).

supabase