The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts scripts/mine_session_logs.py and scripts/score_ideas.py execute the external claude CLI tool to perform LLM-based analysis.
Evidence: Uses subprocess.run(["claude", "-p", ...]) to pass prompts and data to the Claude CLI via standard input.
[DATA_EXFILTRATION]: The skill reads and processes local session history files which may contain sensitive user data or proprietary code snippets.
Evidence: Scans ~/.claude/projects/ for .jsonl session logs and extracts user messages and tool outputs.
Mitigation: The skill implements a PROJECT_ALLOWLIST in mine_session_logs.py to limit the scope of the file scan to specific directories.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface because it interpolates untrusted data from past session logs directly into prompts for the claude CLI.
Ingestion points: Extracted user messages and tool result errors from ~/.claude/projects/*.jsonl (found in scripts/mine_session_logs.py).
Boundary markers: Uses Markdown headers (e.g., ## User Message Samples) as delimiters, which can be bypassed by adversarial content in the logs.
Capability inventory: The skill can execute the claude CLI and write to the local filesystem (YAML backlog files).
Sanitization: Truncates message samples to 200 characters and tool errors to 500 characters, but does not sanitize or escape the content before interpolation.

skill-idea-miner