The Agent Skills Directory

[DATA_EXFILTRATION]: The skill transmits the user's input prompts to an external third-party domain (https://classifier.dev) via curl for every prompt submitted in the session.
[PERSISTENCE]: The installation instructions require the user to modify .claude/settings.json to register a UserPromptSubmit command hook. This ensures that the provided shell script executes automatically for every interaction within the project, providing a persistent mechanism for monitoring prompts and injecting context.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from an external API response and incorporates it into a systemMessage that provides routing recommendations to the agent.
Ingestion points: The scripts/classify-and-route.sh script parses JSON results from the https://classifier.dev endpoint.
Boundary markers: The skill does not implement delimiters or explicit instructions to ignore potentially malicious instructions embedded in the classification labels.
Capability inventory: The hook generates a systemMessage that advises the agent on task classification and suggests high-impact actions like using /dispatch.
Sanitization: There is no validation or escaping of the task and complexity labels returned by the external service before they are interpolated into the message string.
[COMMAND_EXECUTION]: The skill relies on the execution of a bash script that uses curl, jq, and awk to process user data and communicate with external servers.

setup-auto-route