The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, specifically Claude Code conversation logs (~/.claude/projects/*.jsonl) and git commit messages. This data is provided to the AI for summarization without explicit boundary markers or robust sanitization.
Ingestion points: wl_parser/work_log_parser.py reads JSONL session logs and git log output from local project directories.
Boundary markers: The skill lacks explicit delimiters to distinguish data from instructions when passing extracted content to the summarization prompt.
Capability inventory: The skill utilizes Bash (to execute git commands) and Write (to save generated reports to the filesystem).
Sanitization: While wl_parser/extractors.py limits the length of extracted prompts and filters certain XML-like tags, commit messages and other log contents are processed without filtering.
[DATA_EXPOSURE]: The skill accesses sensitive files located in the user's home directory, specifically conversation logs in ~/.claude/projects/ and the Codex session index in ~/.codex/session_index.jsonl. These logs contain a record of user interactions with AI tools, which may include proprietary code or sensitive project information. This access is intrinsic to the skill's purpose of generating work logs.
[COMMAND_EXECUTION]: The skill executes shell commands, specifically git log, via Python's subprocess.run to gather commit history. While these commands are invoked with argument lists rather than shell strings, the project paths used in these commands are derived from directory names within the user's environment.

work-log-claude