The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill clones a remote repository from https://github.com/affaan-m/everything-claude-code.git into a local temporary directory (/tmp).
[REMOTE_CODE_EXECUTION]: Content from the unverified repository is copied directly into active skill directories (~/.claude/skills/ or .claude/skills/). These files contain instructions that are subsequently loaded and executed by the AI agent.
[COMMAND_EXECUTION]: The skill executes multiple shell commands including 'rm', 'git clone', 'mkdir', 'cp', and 'grep' to perform file system operations and environment configuration.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by ingesting untrusted markdown data from a remote repository. It performs 'optimization' (reading and editing) on these files without sanitization or boundary markers while maintaining write access to the agent's capability directories. Evidence Chain:
Ingestion points: Cloned repository contents via 'git clone' in SKILL.md.
Boundary markers: Absent; the skill processes external markdown files directly.
Capability inventory: File system write access ('cp', 'mkdir'), file deletion ('rm'), and text searching ('grep') across SKILL.md.
Sanitization: Absent; the skill reads and modifies untrusted markdown files in-place during the optimization step.

configure-ecc