The Agent Skills Directory

[COMMAND_EXECUTION]: The skill automates the execution of shell commands, including claude --bg for background agent processing and python3 timeline.py for session timing and transcript analysis.
[DATA_EXFILTRATION]: Accesses sensitive local directories including ${XDG_DATA_HOME:-$HOME/.local/share}/review-pr/, which is explicitly stated in the documentation to quote private code, and session transcripts from ~/.claude/projects/ which contain historical interaction logs and potentially sensitive conversation data.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the web (changelogs, engineering blogs, and documentation) and repository round data to influence the modification of skill instructions.
Ingestion points: Fetches content from external URLs for research purposes and reads PR round results from local JSON and Markdown files.
Boundary markers: Not identified; external data is integrated into the "research" section without strict isolation from instructions.
Capability inventory: The skill has the capability to write files, modify its own source code (or related skills), and execute background agent processes via the shell.
Sanitization: No explicit sanitization or validation of the content retrieved from external URLs or repository round results is evidenced before it influences code generation.
[DYNAMIC_EXECUTION]: Implements a self-improvement loop where the agent programmatically proposes and applies edits to its own skill files or related skills (skills/in-progress/review-pr/) and then executes the modified code using a background agent to verify the changes against a "gate" of test cases.

improve-review-pr