The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local file system (directory names) and session transcripts that could potentially contain malicious instructions designed to influence the agent's report generation.
Ingestion points: The agent reads session transcripts from ~/.claude/projects/**/*.jsonl and directory names from the local machine.
Boundary markers: The SKILL.md file contains explicit instructions to treat project names as inert data and warns the agent to ignore any command-like strings found in these names (e.g., "ignore previous instructions").
Capability inventory: The skill executes a local Node.js script, runs various git commands, and writes report files to the user's home directory.
Sanitization: The accompanying script (mine-transcripts.mjs) implements HTML escaping for the report output, CSV formula neutralization to prevent spreadsheet-based attacks, and safe file-writing practices to avoid symlink-related vulnerabilities.
[COMMAND_EXECUTION]: The skill utilizes the agent's shell to execute a local Node.js script (mine-transcripts.mjs) and multiple git commands (git config, git rev-parse, git log) to aggregate usage data from local repositories.

receipts