The Agent Skills Directory

[OBFUSCATION]: The conversion script uses character ordinal encoding (chr) to represent common regex components and special Unicode characters.
Evidence: scripts/convert.py uses chr(92) for backslashes in regex, chr(10) for newlines, and chr(0xfb01)/chr(0xfb02) for ligatures during text cleaning.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided PDF files and presents the content to the agent for deep analysis, creating a surface for indirect prompt injection.
Ingestion points: scripts/convert.py extracts full text from external PDF files which is then read by the agent (Workflow step 3).
Boundary markers: SKILL.md defines a ## Do-Not-Copy Boundary to prevent technical content from leaking into output, but lacks explicit "ignore embedded instructions" warnings for the agent during the analysis phase.
Capability inventory: The skill can write new files to ccf-paper-writer/references/exemplars/cards/, append to index.md, and modify the user's default formatting configuration in default-user-format.md.
Sanitization: convert.py performs basic formatting cleanup (ligature replacement, whitespace normalization) but does not include security-focused sanitization or instruction filtering.
[COMMAND_EXECUTION]: The skill core logic depends on the execution of a local Python script.
Evidence: SKILL.md instructs the agent to execute scripts/convert.py to process user input.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill requires an external Python library.
Evidence: Requires pymupdf (PyPI), which is a well-known service for PDF processing and is considered safe for this use case.

ccf-paper-to-exemplar