@freak30

caveman-xiaoz applies aggressive token compression to responses by stripping articles, fillers, hedging, and pleasantries while preserving technical substance, code, and ordering. three triggerable modes (lite/full/ultra) and sub-skills for commits and code review.

skill assists with academic paper refinement using a three-pass methodology targeting rhetorical density reduction, evidence-level tagging, and explicit caveat disclosure for sci submissions. outputs include annotated manuscript, modification rationale, and ai-trace assessment.

skill-router reduces routing overhead in large skill graphs by bucketing scenarios and activating only top-3 relevant skills per request, with historical failure tracking to downweight ineffective skills.

ai-consciousness-core proposes five modules for confidence detection, event logging, prediction verification, identity anchoring, and incremental distillation. api contracts are sketched but lack implementation detail, error handling, and operational validation.

idle-learning schedules hourly scrapes of github trending and arxiv papers, logs findings to json, and runs a distillation job at 03:00 to write summaries into long-term memory. architectural intent is clear but operational details are absent.

长视频（YouTube链接/本地视频）→ N个短高光片段+烧字幕+AI封面。100%免费方案：yt-dlp字幕提取+MiniMax LLM选高光+MiniMax图片生成封面+ffmpeg剪辑。

小Z的元认知模块——"知道自己知道什么"的置信度检测。当小Z判断自己的回答置信度低于阈值时，在回复前加"[🤔 不确定]"标记。研究来源：UQLM (arXiv:2602.17431) + Hallucination Signals (arXiv:2604.13068)