The Agent Skills Directory

[SAFE]: The skill utilizes a bundled Python script (scripts/fill_icmje_coi.py) that relies exclusively on the Python standard library (zipfile, argparse, json, re). It does not download or execute unverifiable external dependencies.
[SAFE]: Input sanitization is implemented in the _replace_in_zip function, which explicitly escapes XML special characters (&, <, >) before performing string substitution, preventing XML injection attacks.
[SAFE]: The skill follows a human-in-the-loop design with three distinct 'Gates' requiring user approval before generation, verification, and circulation of the generated documents.
[SAFE]: The provided Word template (templates/icmje_coi_seed_synthetic.docx) is documented as a PII-scrubbed synthetic seed. The instructions explicitly forbid the use or commitment of seeds containing real author data to prevent accidental data exposure.
[SAFE]: Verification logic is provided via a static Python snippet in the instructions to ensure the integrity of the generated files (e.g., checking checkbox counts and confirming the absence of seed placeholders).
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input such as manuscript titles and author names.
Ingestion points: Data enters through Phase 1 intake (Title, Date, Author List) in SKILL.md.
Boundary markers: The skill requires an intake summary presentation (Gate 1) for user approval before processing.
Capability inventory: The bundled script is limited to file system writes within a user-specified output directory; it lacks network or shell execution capabilities.
Sanitization: The script performs literal string substitution with mandatory XML escaping, neutralizing potential injection payloads targeting the Word document structure.

fill-icmje-coi