The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and act upon data stored in the workspace (summaries, character profiles, world settings). This creates a vulnerability surface where malicious instructions embedded in the novel's text could influence the agent's behavior during memory retrieval.
Ingestion points: The agent is instructed to read files from novel/{书名}/设定/ and summary/chapter_XX.md at the start of every chapter to maintain continuity.
Boundary markers: The instructions do not specify any delimiters or warnings to ignore embedded instructions within the retrieved memory files, increasing the risk that the LLM might follow instructions found in the story text.
Capability inventory: The skill utilizes file read/write operations and the grep shell utility to manage and search data within the workspace.
Sanitization: There is no requirement for the agent to sanitize or validate content before it is re-integrated into its context.

novel-memory