The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data such as video analysis (vlm_analysis.json) and ASR results (asr_result.json). This content is interpolated into prompts for an LLM-based reviewer in scripts/review_response.py. The skill mitigates risks by using structured Markdown headers as boundary markers and truncating input strings using a _clip_text utility.
Ingestion points: vlm_analysis.json, asr_result.json, background_research.json (handled in scripts/review_runner.py).
Boundary markers: The prompt uses specific headers (e.g., ## Timeline evidence, ## 解说草稿) to separate external data from system instructions.
Capability inventory: The skill performs network calls to a specific API and writes process reports to local JSON/Markdown files.
Sanitization: Truncation logic is implemented to prevent prompt overflow or injection via excessively long strings.
[EXTERNAL_DOWNLOADS]: The skill communicates with the Xiaomi MiMo API (api.xiaomimimo.com) for its core AI-driven video understanding and text evaluation features. This is a legitimate dependency on a well-known service provider, and communications are performed securely via HTTPS.
[COMMAND_EXECUTION]: The skill executes local Python scripts (scripts/review.py, scripts/validate.py) for data validation and reporting tasks. These executions are internal to the skill's logic and are not exposed to untrusted user input.
[SAFE]: Credentials (MIMO_API_KEY) are managed via environment variables rather than hardcoded secrets. Furthermore, the skill implements specific redaction logic in scripts/lib.py (_sanitize_api_error) to ensure that API keys and large binary data blobs are not leaked into error logs or diagnostic output.

video-script