The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes an automated content processing pipeline where untrusted text flows between multiple autonomous agents, exposing an execution surface for indirect prompt injection or cross-stage manipulation.
Ingestion Points: Raw user creative requests are captured during the demand collection stage and stored in .sumeru/session/requirements.json and .sumeru/session/user-requirements.json. Subsequently, generated chapter text within chapters/*.md is loaded back into context during the logic review, content polishing, and finalization stages.
Boundary Markers: Absent. The instructions do not prescribe structural delimiters or strict guidance to ensure that down-stream parsing agents separate user-controlled story content from underlying execution logic.
Capability Inventory: The skill orchestrates multiple sub-agents (sumeru-write, sumeru-review, sumeru-polish, sumeru-finalize) that possess permissions to write, overwrite, and review files within the local working directory (e.g., automatic overwriting of files in chapters/ during the review and polish stages).
Sanitization: Absent. No sanitization, content filtering, or validation mechanisms are defined to scan user inputs or generated text for embedded instruction overrides before execution.

sumeru-worldbuilder