The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill processes untrusted session logs which could contain malicious instructions designed to influence the LLM. 1. Ingestion points: Reads .jsonl files from the current directory and ~/.claude/projects/. 2. Boundary markers: No specific delimiters or safety warnings are included for handling the ingested content. 3. Capability inventory: The skill is limited to read-only analysis and generating text prompts; it has no access to shell execution, network requests, or file writing. 4. Sanitization: No sanitization or validation of the log content is performed.
[Data Exposure] (SAFE): Access to agent history files in ~/.claude/projects/ is necessary for the skill's stated purpose and does not target system credentials, private keys, or other high-value secrets.

agent-feedback