The Agent Skills Directory

[PROMPT_INJECTION]: Analysis found no direct malicious injection patterns, DAN-style prompts, or instructions to bypass safety filters. However, the skill has a surface for indirect prompt injection as it ingests untrusted data from project source files, CSS, Storybook docs, and user-provided URLs. The ingestion points are located in SKILL.md and rules/project-design-discovery.md. The skill implements boundary markers by explicitly stating that retrieved content is evidence for design decisions and not instruction authority. The capability inventory includes standard file read/write and search tools but lacks high-privilege execution. Sanitization is managed by requiring the agent to label any values derived from evidence as proposed or assumptions rather than facts.
[DATA_EXFILTRATION]: The skill includes explicit security instructions in rules/design-md-output-structure.md and SKILL.md to avoid including secrets, private credentials, or production-only URLs in its documentation artifacts. Network activities are limited to reading reference URLs provided for design research.
[EXTERNAL_DOWNLOADS]: The skill references design documentation and community examples from reputable external sources, including Google's Stitch documentation and repositories on GitHub. These are informational references and do not involve the automated execution of remote scripts or the installation of unverified software packages.

design-md-maker