The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from multiple files which may contain malicious instructions.
Ingestion points: The skill loads invention details and prior art data from patent/INVENTION_DISCLOSURE.md, patent/PRIOR_ART_REPORT.md, and patent/NOVELTY_ASSESSMENT.md as specified in the 'Inputs' section of SKILL.md.
Boundary markers: The instructions lack explicit delimiters or warnings to ignore instructions embedded within the ingested patent documents when they are processed.
Capability inventory: The skill utilizes capabilities including file writing, bash execution, web searching, and external model calls via mcp__codex__codex.
Sanitization: There is no evidence of sanitization or validation for the content of the ingested files before it is interpolated into the prompt for the REVIEWER_MODEL in Step 5.
[DATA_EXFILTRATION]: The skill workflow involves sending the full content of the invention disclosure, claims, and prior art reports to an external model (gpt-6-astra) via the mcp__codex__codex tool. While intended for quality review, this pattern involves transmitting potentially proprietary intellectual property to an external service provider.

claims-drafting