The Agent Skills Directory

[DATA_EXPOSURE]: The skill accesses local session metadata stored in ~/.claude/usage-data/facets/ to generate friction reports. This is consistent with its stated purpose and does not involve network exfiltration.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from session facets and git history. Ingestion points: ~/.claude/usage-data/facets/*.json, git branch names, and commit messages. Boundary markers: None explicitly defined for prompt interpolation. Capability inventory: Writing Markdown and HTML files to the local file system (BRAINSTORM.md, SPEC.md, report.html). Sanitization: Implements a 'Defensive Parsing Contract' to handle malformed JSON and file errors, though it lacks specific input sanitization for embedded LLM instructions.
[COMMAND_EXECUTION]: No evidence of shell command execution or subprocess spawning. The skill primarily performs file read/write operations and data aggregation.

brainstorm-insights