The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a process for ingesting external project data (JSON and JSONL files) to generate image prompts, which establishes a theoretical attack surface for indirect injection.
Ingestion points: Data enters the agent context from short-drama.json and asset files like 设定集/*.jsonl as defined in the image-prompt-spec.jsonl.md template.
Boundary markers: Instructions specify that generated prompts must be contained within ### 可复制提示词 Markdown blocks to provide clear delimiters.
Capability inventory: Included utility scripts (image_prompt_check.py and selftest.py) are limited to standard library functions for parsing, regex, and file path resolution. They lack network access, administrative shell execution, or file-writing capabilities to arbitrary locations.
Sanitization: The validation scripts implement an active security layer that explicitly blocks internal hashes (sha256), engine-specific control words (e.g., Midjourney syntax), and credential-related fields (e.g., api_key, token) from appearing in the prompt data.

short-drama-image-prompts