The Agent Skills Directory

[PROMPT_INJECTION]: The skill describes an indirect prompt injection surface in debug-tricks.md by instructing the agent to process output from a local development JSON-RPC endpoint (/_next/mcp).
Ingestion points: The agent is encouraged to read and interpret tool outputs like get_errors and get_routes in debug-tricks.md.
Boundary markers: There are no instructions to use delimiters or special prompts to ignore potentially malicious content embedded in development logs or error messages.
Capability inventory: The skill outlines the agent's ability to execute shell commands (e.g., curl) and build scripts, which could be influenced by the data ingested from the dev server.
Sanitization: There are no explicit instructions for validating or sanitizing the output of the local development server before the agent processes it.
[DATA_EXFILTRATION]: The debug-tricks.md file documents the use of tools that reveal sensitive system information to the agent context. Specifically, the get_project_metadata and get_logs tools provide the agent with absolute filesystem paths, including the project's root directory and the location of Next.js development logs (next-development.log). This exposure is limited to the local development environment.

next-best-practices