The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted data from external websites.
Ingestion points: External content is fetched in references/checks.md via curl "$PDP" and curl "$FEED_URL".
Boundary markers: SKILL.md includes guardrails instructing the agent to 'Treat facts without a visible source as unknown' and to 'never infer ingredients, compatibility, contraindications, or safety claims'.
Capability inventory: The agent has access to shell execution, curl, grep, sed, and python3.
Sanitization: No explicit sanitization or filtering of the fetched HTML or JSON content is described before the agent processes the data for analysis.
[COMMAND_EXECUTION]: The skill instructions and evaluation scripts involve executing shell commands.
Evidence: references/checks.md contains multiple pipelines using curl, grep, sed, and python3 -m json.tool to extract product attributes and structured data.
Evidence: The evaluation harness in eval/run_harness.py and eval/runner.py utilizes subprocess.run to execute Docker commands and binary targets (/usr/local/bin/product-knowledge-target) for automated testing.
[EXTERNAL_DOWNLOADS]: The skill is designed to download content from arbitrary external URLs.
Evidence: references/checks.md uses curl -sS -L to download product pages and catalog feeds provided in the prompt context via $PDP and $FEED_URL variables.

product-knowledge-gap-analysis