The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to analyze and modify user-provided manuscript files and project configuration documents across multiple automated pipelines, such as Prose Revision and Novel Development. Sub-agent prompt templates (e.g., in references/prose-revision/phase-1-surface-cleanup.md and references/manuscript-audit/audit-believability.md) instruct the agent to ingest this external content using variables like {{MANUSCRIPT_PATH}} and {{BOOK_PATH}}. The instructions lack explicit boundary directives or safety markers telling the agent to treat manuscript content strictly as data rather than executable instructions. Combined with the agent's access to the Write, MultiEdit, WebSearch, and WebFetch tools, this creates an attack surface where maliciously crafted manuscript content could potentially influence agent behavior or trigger unintended tool use. The skill currently lacks input sanitization or filtering to mitigate this risk.

creative-writing