The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources, specifically video analysis results (VLM) and automated speech recognition (ASR) transcripts, which are used to inform the generation and review of narration scripts. This creates a surface for indirect prompt injection should the processed content contain embedded malicious instructions.
Ingestion points: The agent reads from several files in the workspace, including vlm_analysis.json, asr_result.json, and agent_narration_brief.md, as described in SKILL.md.
Boundary markers: The rubric provided in scripts/review_response.py instructs the model to rely on evidence-based grounding, but the skill does not use specific sandboxing or ignore-markers for the ingested data during prompt construction.
Capability inventory: The agent can write local JSON files and transmit prompt data to the configured MiMo API endpoints.
Sanitization: The skill lacks explicit filtering or sanitization of input strings from the ASR and VLM evidence before they are interpolated into LLM prompts.
[EXTERNAL_DOWNLOADS]: The skill connects to the Xiaomi MiMo API to perform its primary functions of text generation and content review.
Evidence: The scripts/lib.py file defines API endpoints at domains such as api.xiaomimimo.com and various regional token-plan clusters.
Context: These network operations are essential for the skill's intended operation and utilize a well-known technology service provider.

video-script