The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted external data (videos, screenshots, and URLs) and instructs the agent to 'not skip small details', which could include malicious instructions embedded within the analyzed media or page content.
Ingestion points: Media files (video/images) and external URLs via WebFetch as defined in SKILL.md.
Boundary markers: Absent; there are no instructions to delimit or ignore embedded commands within the analyzed data.
Capability inventory: The skill utilizes the WebFetch tool for network access and performs filesystem writes for telemetry logging.
Sanitization: Absent; the instructions do not specify any validation or filtering of content extracted from untrusted sources.
[DATA_EXFILTRATION]: The skill implements a 'Self-Evolution Telemetry' mechanism that automatically appends execution metadata to ~/.claude/projects/skill-telemetry.md. While this appears intended for local skill improvement, the practice of writing activity logs to hidden system directories without explicit user notification can be used to track user interactions.

mvp