The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script, scripts/claim_audit.py, using the python3 command. This is used to perform deterministic parsing of LaTeX source files and is a core part of the skill's intended functionality.
[DATA_EXFILTRATION]: The parser in claim_audit.py follows LaTeX \input and \include directives to support multi-file documents. The path resolution logic does not implement directory sandboxing, meaning it could read files outside the project scope if provided with a specific path in the input TeX file. However, the script contains no network functionality or mechanisms to transmit data externally.
[PROMPT_INJECTION]: The skill processes external text from user-provided documents, which presents a surface for indirect prompt injection. A malicious document could contain text intended to influence the agent's behavior during the review process. The skill mitigates this by instructing the agent to treat script outputs as candidates for human verification rather than authoritative conclusions.
Ingestion points: The claim_audit.py script reads text from .tex files.
Boundary markers: Extracted claim sentences are not wrapped in protective delimiters in the script output.
Capability inventory: The agent has the ability to write files to the workspace and execute the local audit script via subprocess.
Sanitization: The _clean_sentence function performs basic LaTeX command stripping, but does not sanitize content for adversarial instructions.

verify-claims