The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources. * Ingestion points: The hallmark study command in SKILL.md accepts remote URLs. * Boundary markers: Present in references/study.md, which instructs the agent to treat fetched content as untrusted inert data and ignore instructions in HTML, CSS, or comments. * Capability inventory: The agent is authorized to perform network operations (WebFetch) and file writes (design.md, tokens.css, .hallmark/log.json) as described in SKILL.md and references/design-md.md. * Sanitization: Present in references/study.md, which includes a remote URL safety protocol to validate schemes and block private IP ranges.
[DATA_EXFILTRATION]: The skill performs network operations via the WebFetch tool to analyze designs from user-provided URLs. To prevent Server-Side Request Forgery and access to internal data, the skill defines a strict safety check in references/study.md that forbids non-web schemes and private, loopback, or metadata IP addresses.
[COMMAND_EXECUTION]: The skill instructs the agent to manage project-level design tokens and logging. This involves writing to local files such as tokens.css and .hallmark/log.json. Safety instructions in SKILL.md and references/verbs/redesign.md prohibit the deletion of production files or route trees without explicit confirmation.

hallmark