The Agent Skills Directory

[SAFE]: The skill consists exclusively of Markdown-based instructions and JSON evaluation data. It does not include any executable scripts, binaries, or active code components, which inherently limits its risk profile.
[INDIRECT_PROMPT_INJECTION]: The skill functions by processing and editing behavior-defining prose, which is an inherent surface for indirect prompt injection.
Ingestion points: Untrusted prompt artifacts, including SKILL.md, AGENTS.md, and references/*.md files, are the primary inputs for processing.
Boundary markers: The skill explicitly defines boundary markers through its 'Core contract', 'Hard stops', and 'Operating controls', which provide clear instructions on what must remain unchanged.
Capability inventory: The skill's capabilities are restricted to reading and rewriting text-based instruction files.
Sanitization: Risk is mitigated by a mandatory 16-step self-check process ('Required self-check before delivery') and rigorous decision tests designed to detect behavioral drift, qualitative gates, and obligation-strength shifts.
[EXTERNAL_DOWNLOADS]: The README.md provides installation instructions using npx and pnpm that reference the vendor's official GitHub repository (github.com/gohypergiant/agent-skills). These are documented as standard deployment practices for the vendor and do not involve untrusted third-party sources.

accelint-skill-prose