The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content from untrusted external sources (PPTX and HTML files) to perform audits. This architecture creates a surface where instructions embedded within the slide content could potentially influence the agent's behavior.
Ingestion points: The agent analyzes extracted content from .pptx files and reads source .html slide decks.
Boundary markers: The instructions do not explicitly include delimiters or warnings to ignore instructions that might be embedded in the processed slide content.
Capability inventory: The agent can execute local Python scripts (extract_pptx.py, verify_layout.py) and perform local file read/write operations.
Sanitization: The skill does not perform sanitization or validation of the text content extracted from the slides before the agent processes it.
[EXTERNAL_DOWNLOADS]: The skill references standard external dependencies and font resources required for its primary task.
Evidence: Instructions in SKILL.md and scripts/extract_pptx.py mention installing the python-pptx library via the standard pip registry.
Evidence: references/font-discipline.md includes examples of installing fonts such as Noto Serif TC using brew or other system package managers. These are documented as necessary requirements for ensuring font fidelity.
[SAFE]: Analysis of the included Python scripts confirms they are focused on geometric and typographic property extraction. No evidence of network operations, credential harvesting, or unauthorized system modifications was found.

pptx-html-fidelity-audit