The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches a scoring rubric from an external URL (https://jeffbailey.us/blog/2026/01/01/fundamentals-of-software-usability/llm.txt) and treats it as the 'source of truth' for its analysis. If this remote file were compromised, it could influence the agent's behavior via instructions embedded in the rubric.
Ingestion points: SKILL.md (Domain Knowledge section) downloads the file using curl and processes it via awk before the agent reads it.
Boundary markers: None specified for the rubric content; the agent is instructed to let the article 'win' in case of contradictions.
Capability inventory: File writing (docs/usability-review.md), network operations (curl), and browser automation (evaluate_script).
Sanitization: The skill uses awk to extract specific sections (6, 7, and 9), which provides a minor structural filter but does not sanitize the content itself.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to interact with and analyze arbitrary websites provided by the user. This exposes the agent to untrusted data (HTML, text, metadata) that could contain malicious instructions meant to bypass safety filters or influence the report output.
Ingestion points: Browser automation tools, curl for HTML/header fetching, and screenshot analysis.
Boundary markers: Instructions require the agent to record URLs and elements for findings, providing some structure, but no explicit prompt isolation is used for external content.
Capability inventory: File system access for report/screenshot storage and full browser interaction.
Sanitization: None; the agent processes the live rendered state of the target site.
[EXTERNAL_DOWNLOADS]: The skill fetches its core rubric and updates from jeffbailey.us. As this domain is owned by the skill author ('jeffabailey'), this is documented as a vendor resource retrieval rather than an untrusted third-party download.

review-usability