RRe:zZCinzzc123456.hashnode.dev·14h ago · 4 min readBuilding a Knowledge Base from Scratch, EP01An experiment about a habit, not a model In 2026, asking an AI about internal company matters starts with dragging a file into the chat. The move is so smooth that nobody examines it. Before building 00
RRe:zZCinzzc123456.hashnode.dev·2d ago · 8 min readI went directory-diving through DeepSeek Harness's 4,000 plugins, and found the seat nobody tookOn the night of August 13, DeepSeek open-sourced DeepSeek Harness (DSH), an agent substrate built around a radical proposition: everything is a plugin. The model, the tools, the session, the sandbox, 00
RRe:zZCinzzc123456.hashnode.dev·4d ago · 4 min readRerolling is a cost structure, not a luck problem: front-loading keyframes in AI video generationEvery AI video workflow inherits the same lossy step: the same prompt returns a cinematic shot or a broken mess, and industry voices cite 20 pulls for 1 usable take. On per-second billing, every rerol00
RRe:zZCinzzc123456.hashnode.dev·Aug 14 · 6 min readModel selection as a command: intent analysis, candidate recall, and why rankings answer the wrong questionEvery project that calls LLM APIs keeps re-answering the same question: which model for this one. It's not a preference problem — pick too expensive and the unit economics break, pick too weak and qua00
RRe:zZCinzzc123456.hashnode.dev·Aug 12 · 5 min readOne System Prompt, Five Text Jobs: Compliance Screening, Review Tagging and Contract Extraction from the TerminalWhere this started A small-shop owner I know had 300 unread negative reviews and a rejected quote from a review-analytics SaaS. I had a command-line LLM client (Alibaba Cloud Model Studio's bl) and a 00
