The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes company data from external CSV files which may contain untrusted strings, such as company descriptions.
Ingestion points: companies.csv and individual company data fields (description, name) are loaded into the agent's context during the tuning loop and batch scoring.
Boundary markers: The prompt template in Step 3 lacks explicit delimiters or instructions to prevent the sub-agent from following commands potentially embedded in the company description field.
Capability inventory: The skill can execute local scripts via npx tsx and write files to the filesystem (icp-prompt.txt, scored.csv).
Sanitization: No sanitization or escaping of external company data is mentioned before it is interpolated into the evaluator prompt template.
[COMMAND_EXECUTION]: The skill uses local command execution to process data batches.
Evidence: Instructions include running npx tsx to execute a local helper script (scripts/score-batch.ts) for scaling the scoring process.

icp-prompt-builder