The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documentation highlights a /load command intended for dynamic skill loading at runtime. This functionality allows the agent to load and interpret instructions or logic from local file paths, which could be exploited to execute unauthorized actions if the file source is not validated.
[PROMPT_INJECTION]: The system utilizes 'markdown-backed sessions' and a 'markdown-as-database' approach for persistent storage. Ingesting and processing historical data from these files creates a surface for indirect prompt injection, where malicious content embedded in session history or shared files could override the agent's instructions during subsequent loads.
Ingestion points: Markdown session files and dynamically loaded skills referenced in SKILL.md.
Boundary markers: Not specified in the provided instructions; boundaries between data and instructions are managed by the underlying REPL script.
Capability inventory: The skill references file reading, writing, searching, and model switching commands in SKILL.md.
Sanitization: No explicit sanitization or validation of the loaded markdown content is described in the provided text.
[NO_CODE]: The skill package contains operating instructions and metadata but does not include the implementation scripts (such as scripts/claw.js) or executable binary code within the analyzed file, relying instead on external components already present in the environment.

nanoclaw-repl