本文由 AI 分析生成
建立時間： 2026-01-20 來源： https://x.com/XDash/status/2013420707657720175
Summary
A writeup of a “How I AI” (Lenny’s Newsletter) interview with Teresa Torres on moving from GUI tools like Trello to a Claude Code + Obsidian “AI-native” workflow. Her system keeps all tasks as local Markdown files in Obsidian, with a /today slash command that runs Python to scan YAML frontmatter and assemble a daily today.md. It covers automated tagging from a taxonomy, semantic search over notes, a daily arXiv/Scholar research digest focused on Methods and Effect Size, granular context files, and an AI writing partner that fact-checks and enforces her style.
這是對「How I AI」（Lenny’s Newsletter）採訪 Teresa Torres 的整理：她從 Trello 等 GUI 工具轉向 Claude Code + Obsidian 的「AI 原生」工作流。她把所有任務存成 Obsidian 裡的本地 Markdown 檔，用 /today 斜線指令跑 Python 掃描 YAML 前綴，自動彙整出每日的 today.md。內容涵蓋依分類法自動打標籤、筆記語意搜尋、每天聚焦於「研究方法」與「效應值」的 arXiv/Scholar 論文摘要、顆粒化的上下文檔案，以及會事實查核並執行她寫作風格的 AI 寫作夥伴。
Key Points
- Task management lives entirely in local Markdown;
/todayruns Python to aggregate due/overdue/in-progress items into a fresh file. - Auto-tagging from a supplied taxonomy and semantic search make manual filing and exact-keyword recall unnecessary.
- A two-step research digest: AI lists matching papers daily; new PDFs dropped in topic folders trigger deep reads summarizing Methods and Effect Size.
- Context is stored as hundreds of tiny Markdown files indexed via CLAUDE.md, so queries load only relevant profiles — saving tokens and reducing hallucination.
- The AI is a writing partner (real-time fact-check, style reviewer trained on 10 years of her blog), not a ghostwriter;
/clearresets context cheaply since all state is in files.
Insights
The “feed small files, not big files” principle is the transferable lesson: routing queries to scoped context files via CLAUDE.md is a deliberate retrieval design that simultaneously cuts cost and hallucination. Her end-of-conversation prompt — “what did you learn that should go into the context files?” — turns each session into a self-updating knowledge base, a lightweight form of the memory-management problem other practitioners describe.
Connections
Raw Excerpt
她提出了一个非常关键的观点：不要给 AI 喂大文件，要喂小文件。