The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script, scripts/validate_novel_project.py, to verify project structure and JSON schema compliance. This script is limited to reading local files within the project directory and does not perform network operations or require elevated permissions.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data, such as existing manuscripts and user-provided outlines, which constitutes an ingestion point for untrusted content. However, the risk is mitigated by rigorous narrative auditing protocols and boundary markers defined in the 'quality gates' and 'story bible' templates, which ensure that the agent remains focused on narrative consistency and logical validation.
Ingestion points: Reads existing manuscripts and user outlines during project initialization and '续写项目' (continuation) modes.
Boundary markers: Implements strict structured formats using '本章写作约束卡' (writing constraint cards) and JSON ledgers to delimit context.
Capability inventory: Includes local file read/write and execution of a provided Python validation script.
Sanitization: Employs multi-stage revision protocols (described in references/revision-protocol.md) that require logical justification and evidence for every change.

novel-orchestrator