The Agent Skills Directory

[DATA_EXFILTRATION]: The skill performs broad automated scans for sensitive information across the user's home directory and project workspace.
Technical evidence: The scripts/scan_setup.py script targets sensitive paths including ~/.ssh, ~/.aws, ~/.azure, and .env files to check for the presence of credentials.
In high-intensity modes (ultra, yolo), the skill mines historical interaction transcripts located at ~/.claude/projects/*.jsonl. This exposes the user's entire history of agent interactions to the current agent's context.
Although the script logic specifically avoids capturing raw secret values in its output JSON, the metadata (file paths, line numbers, and rule matches) provides the agent with a detailed map of the user's security configuration and sensitive data locations.
[INDIRECT_PROMPT_INJECTION]: The skill ingests a large volume of untrusted data from the local environment which is then processed by the agent.
Ingestion points: The agent parses findings from the scan script which includes data from local CLAUDE.md, AGENTS.md, and project-specific rule files. It also reads sample lines from historical transcripts in yolo mode.
Boundary markers: There are no explicit instructions in the skill to use unique delimiters or ignore potentially malicious instructions embedded in the historical transcripts or project files being audited.
Capability inventory: The skill has significant write capabilities, including modifying .claude/settings.json, creating scratchpad files, managing git repositories, and recommending/applying security and workflow fixes.
Sanitization: The audit process lacks a sanitization layer for the text content read from transcripts and external configuration files, relying on the model to differentiate between data and instructions.
[COMMAND_EXECUTION]: The skill executes multiple system commands via a Python subprocess to gather environment metadata.
Tools executed include git (for repository status), lsof (to detect exposed network listeners), ps (to identify running autonomous agents), tailscale (to check VPN status), and sysctl/systemd-detect-virt (for virtualization detection).
The logic uses the subprocess.run list-based interface, which is a standard practice for local system audits but grants the agent detailed insight into the host's running state and network topology.

improve-my-agent-setup