The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it ingests untrusted user data (manuscript titles and author names) to populate generated documents.
Ingestion points: Manuscript title, submission date, and author lists are provided by the user or extracted from the conversation in Phase 1 of the SKILL.md instructions.
Boundary markers: The skill does not use explicit boundary markers for the data, instead performing direct string replacement within the document's XML structure.
Capability inventory: The skill utilizes file writing and bash command execution tools to generate, organize, and verify the resulting .docx files.
Sanitization: The bundled script scripts/fill_icmje_coi.py includes proactive sanitization by escaping XML special characters (&, <, >) in user-provided strings before they are interpolated into the document, effectively preventing XML injection attacks.
[EXTERNAL_DOWNLOADS]: The skill provides references to official ICMJE templates and guidance from icmje.org as part of its documentation and seed provenance description. No automatic downloads or remote code executions are performed by the skill's scripts.
[COMMAND_EXECUTION]: The skill instructions include standard bash commands for file verification and batch processing. These commands are executed locally within the agent's environment to ensure document integrity and are not exposed to unvalidated user input.

fill-icmje-coi