The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the data files it is designed to analyze.
Ingestion points: The skill reads contents from .tex source files and raw results in .json, .csv, .tsv, and .yaml formats from the local filesystem (e.g., paper/ and results/ directories).
Boundary markers: The prompt used for the gpt-5.4 reviewer lacks explicit boundary markers or instructions to treat the ingested file content as untrusted data, increasing the risk that embedded instructions within a paper or data file could influence the audit outcome.
Capability inventory: The skill utilizes Bash(*), Read, Write, Edit, and mcp__codex__codex. The resulting audit report (PAPER_CLAIM_AUDIT.json) is designed to be consumed by other skills like auto-paper-improvement-loop, which may perform automated edits based on the audit findings.
Sanitization: The skill does not perform any sanitization, escaping, or validation of the file content before interpolating it into the reviewer prompt.

paper-claim-audit