The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute a local Python script (scripts/benchmark_fetch.py) to perform search, cost estimation, and data retrieval tasks. It also mentions the use of whisper-cpp for audio-to-text transcription.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze high volumes of external, user-generated content including video titles, cover text, and transcripts. This creates an attack surface where malicious instructions embedded in social media posts could attempt to override the agent's logic during the teardown or roadmap generation phases.
Ingestion points: Data fetched from Douyin, Xiaohongshu, and WeChat Channels (via scripts/benchmark_fetch.py), and transcriptions generated from video content.
Boundary markers: The instructions do not define specific delimiters or sanitization protocols for the extracted text when it is re-interpolated into the analysis prompt.
Capability inventory: File read/write, local script execution, and network access via API.
Sanitization: No explicit sanitization or filtering of external content is described before analysis.
[EXTERNAL_DOWNLOADS]: The skill interacts with the tikhub.io API to retrieve social media metadata and content. It also recommends the installation of the whisper-cpp utility for users requiring automated transcription capabilities.
[CREDENTIALS_SAFE]: The skill guides the user to store their TIKHUB_API_KEY in a .env file (~/.env), which is documented as a safe and standard practice for secret management in CLI-based tools.

aikey-benchmark