The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted communication data (mbox, eml, Slack, WhatsApp) to generate writing samples.
Ingestion points: The scripts/build_corpus.py script reads external files provided as arguments to create a normalized corpus.
Boundary markers: The script does not inject markers into the corpus.jsonl output, but SKILL.md provides explicit warnings to the agent against reading the entire corpus into context to prevent injection.
Capability inventory: Script capabilities are restricted to local file I/O and directory management with private access controls (0600/0700 permissions).
Sanitization: The builder implements robust sanitization, including PII redaction (email, phone, URL, names), HTML stripping, and removal of potentially problematic Unicode directionality marks.
[COMMAND_EXECUTION]: The scripts/test_build_corpus.py script uses subprocess.run for its self-test suite.
The execution is limited to invoking the local build_corpus.py script using the current system's Python executable for validation purposes.
[EXTERNAL_DOWNLOADS]: The references/export-guide.md documentation provides users with links to community tools for exporting data.
These are informational links (e.g., to GitHub repositories for export utilities) for manual user intervention and are not accessed or executed by the skill.

voice-corpus