The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including user-provided thesis drafts, similarity detection reports, and school formatting samples.
Ingestion points: Untrusted data enters the agent context through the existing thesis draft, project evidence files, and similarity/AIGC PDF reports specified in SKILL.md.
Boundary markers: The skill lacks explicit prompt-level delimiters or 'ignore instructions' warnings for external text, though it mandates that the repository remains the 'source of truth' to mitigate hallucination or external influence.
Capability inventory: The skill has the capability to execute local scripts, perform browser automation via Playwright, and write to local DOCX files.
Sanitization: No explicit sanitization or filtering of external document content is documented before processing.
[COMMAND_EXECUTION]: The skill includes two Python scripts, scripts/chapter_quality_check.py and scripts/final_evaluation.py, used to evaluate manuscript quality.
These scripts perform deterministic text analysis, word count verification, and JSON processing.
The scripts do not perform network operations, access sensitive system paths, or utilize privileged commands.
[EXTERNAL_DOWNLOADS]: The skill instructions in SKILL.md and references/docx-draft-salvage-and-runtime-screenshot-playbook.md describe using Playwright for capturing runtime screenshots.
Playwright is a well-known browser automation library and service.
The skill uses it to verify implementation evidence from the user's project runtime, which is a standard functional requirement for the skill's stated purpose.

academic-paper-composer