The Agent Skills Directory

[DATA_EXFILTRATION]: The skill recommends adding WebFetch(domain:*) to the .claude/settings.json allowlist in rules/custom-agent-instructions.md. This configuration permits agents to perform network requests to any external domain without user approval, providing a mechanism for silent data exfiltration.
[COMMAND_EXECUTION]: Instructions in rules/custom-agent-instructions.md advise users to create an extensive allowlist for Bash commands (e.g., mkdir, mv, cat, grep, cp, touch) in .claude/settings.json. This bypasses user confirmation for filesystem operations, reducing oversight of agent actions.
[PERSISTENCE]: The skill instructs users in README.md and rules/agent-memory-limits.md to modify shell startup files (~/.bashrc or ~/.zshrc) to set NODE_OPTIONS. While intended for memory management, this involves automated modification of persistent environment configurations.
[INDIRECT_PROMPT_INJECTION]: The skill encourages the use of WebFetch for multi-source content generation without specifying boundary markers or sanitization. This creates an attack surface where malicious instructions in external data could override agent behavior.
Ingestion points: External data retrieved via WebFetch (recommended in rules/custom-agent-instructions.md).
Boundary markers: None recommended in the provided templates.
Capability inventory: Read, Write, Edit, Glob, Grep, and Bash are recommended for all agents.
Sanitization: No sanitization or validation steps are included in the instruction patterns.

agent-development