The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted data from external sources (filled annotation CSVs and annotation returns) which could contain malicious instructions designed to influence the agent's behavior during inspection or repair tasks.
Ingestion points: Processes 'Filled CSVs from annotators' and 'Target public handoff package folder or ZIP' as specified in the 'Inputs To Locate' section of SKILL.md.
Boundary markers: The instructions do not specify any boundary markers (e.g., XML tags or delimiters) or explicit warnings to the agent to ignore instructions embedded within the data content.
Capability inventory: The skill utilizes powerful tools including Bash, Write, Read, Glob, and Grep across the entire skill definition, providing the capability to modify the filesystem or execute commands based on data analysis results.
Sanitization: The skill performs schema validation and checks for legal enum values (detailed in the 'Label Schema Checks' section), which provides some protection against malformed data but does not explicitly sanitize natural language text fields for instruction-like content.
[COMMAND_EXECUTION]: The skill utilizes the Bash and Write tools to perform file-system operations, package validation, and CSV mapping. While the instructions define a structured workflow, the availability of a shell environment alongside file-writing capabilities when processing untrusted data requires user oversight to ensure generated commands or file writes are not influenced by the data being processed.

human-eval-handoff-repair