YouTube Data API engineering, research intelligence, knowledge management, data collection, ETL pipelines, automation engineering, information architecture, documentation systems, quality assurance, and agentic workflow design. Specializes in transforming YouTube content sources into structured, AI-ready research assets for NotebookLM and modern AI knowledge workflows.
Open-source CLI (MIT, 2k+ GitHub stars, Hacker News front page). Turn any video URL or local file into scene-aware keyframes plus a timestamped, speaker-labeled transcript your LLM can actually read. - Runs 100% on your machine, nothing is uploaded - Works with Claude, ChatGPT, Gemini or local models - URL videos reuse the platform's own captions, Whisper otherwise New in 0.10.5: Apple Silicon GPU transcription, about 6x faster on M-series Macs, same no-fake-captions gate.
Stop confidently-wrong LLM answers. This instruction set forces calibration: every uncertain claim carries a confidence value, every proposed fix names how it breaks and what undoing it costs, every constraint you set is honored or questioned—never silently dropped. Claims are typed, revisions tracked, contradictions logged, drift flagged. Audit and checkpoint on command. Correctness outranks brevity, unconditionally. For technical work where a plausible wrong answer is the expensive one.
Daily arXiv reading across NLP, computer vision, RL, and LLM systems through ICML, NeurIPS, and ICLR cycles. This skill forces a judgment — contributions, results vs. prior SOTA, honest limitations, and reproduction difficulty — so you decide read/skip/implement in under 5 minutes. Covers LLM architecture, MoE scaling, efficient attention, RLHF, and vision-language models.
Structure project docs as modular memory + reasoning files so AI agents load only what they need — cutting token usage by 70–96% per query. Grounded in Anthropic, LangChain, and Crawl4AI's production llm.txt research. Implements write/select/compress/isolate context engineering strategies.
LLMエージェントがセッションをまたいで記憶を保持するための永続メモリ層。抽象基底クラス MemoryStore ＋ LocalJSONStore 実装で、保存・取得・検索を再利用可能に。Karpathy の「LLM OS（記憶＝ディスク層）」着想。バックエンド差し替え可・Python標準ライブラリのみ・LLM呼び出しゼロ・ユニットテスト同梱。
自分が権利を持つ既存noteを診断し、無料記事・有料記事・販売導線・企画再構築のどれが適切かを判断。タイトル、構成、無料・有料ライン、CTAを再設計し、完成稿とX告知まで作ります。
장편은 분량이 길어질수록 초반의 설정과 정보가 후반까지 일관되게 이어지는지, 인물의 변화와 사건의 인과관계가 충분히 축적되는지, 중반부의 반복과 서사 정체가 발생하지 않는지를 함께 확인하는 것이 중요합니다. LUNIR 루니르는 소설 원고를 구조화하고, 미리 설계된 분석 절차와 평가 기준에 따라 여러 단계로 검토한 뒤 결과를 교차 확인하는 소설 원고 전용 정밀분석 프로그램입니다. 플롯·인물·개연성·시점·장면·대화·문체·서사 속도·설정 연속성· 복선과 회수·클라이맥스·결말 등 30개 영역을 원고 전체 흐름에서 종합 분석하고, 주요 문제와 근거·수정 방향·수정 우선순위를 보고서로 제공합니다. [구매 전 확인] 120,001~250,000자 장편 소설 원고 전용 · 공백·줄바꿈 포함 기준
讓 Claude Desktop 每次跳出「需要人工批准」的請求，即時推播到手機。你可以在手機上按 Allow/Deny、從多個選項挑一個，或直接打字回覆，決定會即時回灌到桌面——不必守在電腦前。僅支援 Windows + Claude Desktop，需搭配免費開源的 ntfy App 收推播。安裝精靈會自動產生你專屬的加密 token；開機自啟動與實機驗收需要你在場操作一次完成。（選配 Tailscale，供行動網路開啟多選/打字面板時，面板服務會對區網或 tailnet 開放以便手機連線）