The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content from external websites retrieved via the host's WebFetch tool. This creates a surface for indirect prompt injection where malicious instructions could be embedded in fetched pages to influence the agent.
Ingestion points: Extracted content from WebFetch is used in the evidence gathering and verification stages described in SKILL.md and scripts/prompts.py.
Boundary markers: The skill uses structured JSON schemas and templates to isolate data, although it does not include explicit 'ignore instructions' delimiters for the fetched content itself.
Capability inventory: The skill utilizes WebSearch, WebFetch, and the execution of bundled Python scripts (scripts/*.py).
Sanitization: The skill implements an 'adversarial quorum' mechanism requiring three independent voters to evaluate a claim. This requirement for independent contexts acts as a strong architectural mitigation against poisoning from individual external sources.
[COMMAND_EXECUTION]: The skill relies on executing bundled Python scripts (dedup.py, factcheck.py, rank.py, prompts.py) to perform ranking, deduplication, and quorum logic. These scripts are self-contained and use only the Python standard library. A test script (scripts/test_factcheck.py) also uses subprocess.run to validate CLI behavior of the bundled tools.
[EXTERNAL_DOWNLOADS]: The skill gathers factual evidence by performing web searches and fetching page content using the host agent's native tools for the intended purpose of point-checking claims.

fact-check