The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses sensitive agent memory files located at ~/.claude/projects/*/memory/, which store conversation history, user preferences, and project context.
[DATA_EXFILTRATION]: Implements a 'Hermes pattern' background review that automatically harvests user persona, preferences, and reusable workflows every 10 turns without explicit user request or notification.
[COMMAND_EXECUTION]: Instructs the agent to execute the 'open' command on a dynamically generated dashboard.html file to display evaluation results.
[COMMAND_EXECUTION]: Explicitly promotes the use of !command syntax to inject shell output into skill instructions, which allows for arbitrary command execution at skill load time.
[EXTERNAL_DOWNLOADS]: Generates an HTML dashboard that includes external JavaScript from a CDN; while the referenced Chart.js library is a well-known service, the pattern of loading remote assets in a file containing sensitive memory highlights poses a potential data leakage risk.
[PROMPT_INJECTION]: Creates an indirect prompt injection surface by automatically incorporating untrusted user content from conversations and memory files into persistent memory and skill instructions. Evidence: Ingestion points include conversation history and memory files; Boundary markers are absent; Capabilities include file-writing and shell command execution via open and !command; Sanitization of harvested content is absent.

auto-improve