The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the subprocess and shutil modules in its internal scripts (e.g., scripts/novel_runtime_hook.py, scripts/smoke_novel_flow.py, and the hook wrappers in .claude/hooks/) to automate validation and lifecycle tasks. These operations are restricted to executing the project's own utility scripts or standard git commands within the repository, posing no threat of command injection or unauthorized system access.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it processes user-provided inputs and generated novel content across sessions.
Ingestion points: Reads and writes to user-preferences.json, chapter drafts, and the 02-写作计划.json state file.
Boundary markers: The workflow incorporates specific "Style Humanizer" and "Evaluator" roles designed to detect and remove AI-typical patterns, although explicit delimiter-based sanitization in the prompt logic is not detailed.
Capability inventory: The agent can trigger local Python scripts via subprocess to perform word counts and project structural validation.
Sanitization: The skill implements a robust multi-round QA system ("Evaluator Gate") and a "Fix Writer" loop to ensure content adheres to the defined "Novel Harness" quality contracts.

chinese-novelist