The Agent Skills Directory

[SAFE]: The skill implements its logic through bundled Python scripts (scripts/lint_consistency.py and scripts/lint_figure_locale.py) that use regex for deterministic text analysis, avoiding risky dynamic evaluation.
[COMMAND_EXECUTION]: Subprocess execution is confined to internal testing scripts (tests/test_consistency_controls.py) used to verify the linter's output against synthetic fixtures.
[DATA_EXFILTRATION]: While the skill reads manuscript files and figure sources, it contains no network communication code, preventing data from being transmitted outside the local environment.
[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted manuscript input but mitigates injection risks through defined workflow gates. Ingestion points: Manuscript markdown and figure sources (scripts/XML). Boundary markers: Explicit instructions to use linter output as the source of truth and a 'Phase 2' user gate. Capability inventory: Read/Write/Edit and local Bash for running the linter. Sanitization: The linter provides deterministic reports, and the agent is instructed never to apply changes without human review.

polish-language