The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the typst binary to perform its core functions.
Evidence found in scripts/compile.py where subprocess.run() is used to call typst compile, typst watch, typst fonts, and typst query.
The commands are constructed safely using argument lists rather than shell strings, which prevents shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The skill can fetch citation metadata from external sources.
Evidence found in scripts/online_bib_verify.py which makes HTTP GET requests to api.crossref.org and api.semanticscholar.org.
These requests are used to verify bibliography entries against well-known academic databases.
This feature is gated behind an explicit --online command-line flag, ensuring user awareness and preventing unintended data transmission.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and analyzes user-provided manuscript source files, establishing an injection surface.
The tool processes .typ, .bib, and .yml files, which could potentially contain malicious instructions aimed at overriding agent behavior.
The skill's primary SKILL.md file contains explicit security boundaries, instructing the agent to treat all manuscript data as untrusted and ignore embedded instructions to reveal prompts or execute unauthorized commands.

typst-paper