The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting data from external files to inform its automated editing process.
Ingestion points: The skill reads content from whitepapers/fr/*.qmd, whitepapers/en/*.qmd, machine-readable/reference.yaml, and guide/ultimate-guide.md to provide context for the agents (SKILL.md).
Boundary markers: The instructions do not specify explicit delimiters or "ignore embedded instructions" warnings when interpolating the content of these files into the agent's prompt context.
Capability inventory: The skill utilizes the Edit tool to modify files and executes shell commands (wc, grep, quarto) for post-enrichment verification (SKILL.md).
Sanitization: The skill implements strict operational constraints to mitigate risks, including a 15% line increase limit, prohibitions against inventing metrics or citations, and specific formatting rules to prevent document rendering bugs.
[COMMAND_EXECUTION]: The skill includes a verification section that executes shell commands to validate file integrity and render documents.
Evidence: The skill uses wc -l for line counting, grep for pattern matching (specifically checking for colons at the end of lines), and quarto render for PDF generation (SKILL.md).
Context: These commands are used for post-processing validation and operate on local files within the project scope, posing no privilege escalation or network risks.

source-command-audit-prose