知識萃取主控 PromptKnowledge extraction master prompt
用 B／V 代碼啟動文件清理與核心概念萃取。Use B/V codes to trigger document cleanup and core concept extraction.
知識萃取入門
Knowledge Extraction Guide
把工作經驗變成 AI 能用的知識系統 Turn work experience into an AI knowledge system
這不是把資料收進硬碟，而是把現場判斷、提問方式、文件方法與輸出規則整理成可重複使用的工作流程。 This is not file storage. It turns judgment, questioning patterns, document methods, and output rules into a repeatable workflow.
知道哪些資料值得留下
Know what is worth capturing
把摘要升級成可用方法
Move beyond summaries into methods
用 Skill 與本機工作流延續成果
Use Skills and local workflows to keep improving
錄音 → 清理 → 萃取 → Skill → 本機工作流
Recording → cleanup → extraction → Skill → local workflow
每一頁先看圖，再決定要不要展開Read the diagram first, then open the depth you need
主畫面維持簡報式節奏；細節放在可折疊區塊，讓第一次閱讀不被長文打斷，也讓想實作的人有足夠例子可照做。The main view stays slide-like. Details live in expandable blocks, so first-time readers keep momentum while builders still get practical examples.
三分鐘掃描Three-minute scan
只看每頁標題、圖示與黑底原則，先掌握整套方法的順序。Use titles, diagrams, and rule blocks to understand the full sequence.
十分鐘理解Ten-minute read
展開「怎麼判斷」與「常見錯誤」，理解為什麼不能只做摘要。Open the judgment and mistake panels to see why summaries are not enough.
三十分鐘實作Thirty-minute build
照著練習與下載材料，把一份逐字稿或文件跑成可用的知識卡。Use the exercises and downloads to turn one transcript or document into reusable knowledge cards.
01
先留下現場，不急著整理Capture first. Organize later.
很多專業知識不在正式報告裡，而是在追問、停頓、改口、反應與臨場判斷裡。Much of the real knowledge is not in the final report. It lives in questions, pauses, reframing, reactions, and judgment calls.
先錄音、轉錄、保存對話，再決定要萃取什麼。不要只留下會議摘要。Record, transcribe, and preserve the conversation before deciding what to extract. Do not keep only a meeting summary.
原則：先保留脈絡，再整理結論。Rule: preserve context before compressing conclusions.
WHY為什麼摘要不夠Why a summary is not enough
摘要通常只留下結論，但專業判斷常藏在結論前面的鋪陳：你先接住對方哪一句話、在哪裡停下來追問、什麼時候把話題從意見拉回事實。A summary keeps conclusions, but expertise often appears before the conclusion: what you noticed, where you paused, and how you moved the discussion from opinion back to evidence.
- 要留下：原始問題、追問、對方反應、轉折點、最後判斷。Keep: original issue, follow-up question, response, turning point, final judgment.
- 先不要刪：停頓、改口、反覆說明與看似零散的補充。Do not delete too early: pauses, reframing, repeated explanations, and scattered additions.
CASE實例：一場輔導會議要留下什麼Example: what to keep from one coaching session
如果團隊一直談技術規格，真正要保存的不是「技術很好」這個結論，而是顧問如何把問題改問成：誰會付錢、現在用什麼替代做法、第一階段要切哪個場域。If a team keeps talking about technical specs, keep how the consultant reframed the issue: who pays, what workaround exists today, and which first segment should be chosen.
可用欄位：會議目的｜對象狀態｜顧問追問｜對方反應｜轉折點｜可重用手法。Suggested fields: meeting purpose | audience state | consultant question | response | turning point | reusable method.
現場訊號不是雜訊Signals from the field
清理不是潤稿，是保留可分析的文本Cleanup makes text analyzable
02
逐字稿不是摘要A transcript is not a summary
逐字稿要先變成乾淨文本：去掉辨識錯誤、錯字、無意義字幕、斷句與多餘標點。A transcript must first become clean text: remove recognition errors, typos, stray captions, broken sentences, and bad punctuation.
第一輪只清理，不急著詮釋；第二輪才做情境、方法與範例萃取。The first pass cleans. The second pass extracts context, methods, and examples.
如果一開始就摘要，最有價值的互動細節會先被壓掉。If you summarize too early, the most useful interaction details disappear first.
2 PASS兩輪處理法The two-pass method
- 第一輪只處理文本品質：錯字、斷句、重複字幕、講者切換、標點與段落。Pass 1 improves text quality only: typos, sentence breaks, repeated captions, speaker turns, punctuation, and paragraphs.
- 第二輪才做知識萃取：情境、問題階段、方法選擇、判斷依據、few-shot 片段。Pass 2 extracts knowledge: context, problem stage, method choice, judgment criteria, and few-shot clips.
CHECK清理時要避免的事What cleanup should avoid
不要讓 AI 在清理階段幫你改寫成漂亮文章。那會讓現場用語、卡住的地方、顧問的追問節奏一起消失。Do not let AI turn cleanup into polished writing. That removes the field language, the stuck points, and the rhythm of questioning.
清理後的好文本，應該讀得順，但仍看得出這是一場真實對話。A clean transcript should read smoothly while still feeling like a real conversation.
03
流動知識：現場怎麼推進Flow knowledge: how the work moves
課程、會議、諮詢與輔導都會產生流動知識。重點不是「講了什麼」，而是「當時怎麼問、怎麼判斷、怎麼推進」。Classes, meetings, consulting, and coaching create flow knowledge. The point is not only what was said, but how the work moved forward.
把對象、產業、問題階段、對方心態、限制條件與提問手法標記出來，AI 才能學到工作風格。Mark the audience, industry, problem stage, mindset, constraints, and questioning pattern so AI can learn the working style.
FIELDS流動知識的標記欄位Fields for flow knowledge
- 對象：創辦人、經理人、學員、客戶、團隊成員。Audience: founder, manager, learner, client, team member.
- 情境：課程、會議、諮詢、輔導、簡報修正、會後交付。Setting: class, meeting, consulting, coaching, deck revision, follow-up delivery.
- 判斷：對方是防衛、表面開放、方法不足，還是已經能共創。Judgment: defensive, politely open, lacking method, or ready to co-create.
- 手法：接住、追問、重構、要求證據、安排下一次交付。Method: acknowledge, probe, reframe, ask for evidence, define next delivery.
SHOT把對話轉成 few-shotTurn dialogue into few-shot examples
few-shot 不是漂亮範例，而是讓 AI 看見「遇到這種回答時，下一句該怎麼接」。每段只要保留 4 到 8 句，重點是轉折清楚。A few-shot is not a polished sample. It teaches AI what to say next when it sees a certain response. Keep 4 to 8 lines, with a clear turning point.
格式：情境一句話｜對方原話｜顧問追問｜對方反應｜手法名稱｜下次可重用條件。Format: context sentence | client quote | consultant question | response | method name | reuse condition.
從互動到 few-shot 範例From interaction to few-shot examples
文件要轉成架構，不是堆成全文Documents become structure, not piles of text
04
靜態知識：文件怎麼變成方法庫Static knowledge: documents become method libraries
書籍、講義、投影片與標準文件，不應只被塞進全文資料庫。它們要先清成 Markdown，再重建章節、概念、關係與方法。Books, handouts, slides, and standards should not be dumped into a full-text pile. Clean them into Markdown, then rebuild chapters, concepts, relationships, and methods.
圖、表、流程與架構不能丟掉；需要改寫成能被搜尋與引用的文字說明。Charts, tables, flows, and frameworks should not be discarded. Rewrite them as searchable, citable explanations.
OUTPUT靜態知識應該輸出什麼What static extraction should produce
- 知識架構圖：每一章在解決什麼問題，章節之間如何推進。Knowledge map: what each chapter solves and how chapters progress.
- 概念卡：名詞定義、適用條件、不適用情境、常見誤用。Concept cards: definition, use cases, non-use cases, common misuse.
- 方法流程：把書中的框架改寫成顧問或教學能用的步驟。Method flow: turn frameworks into steps usable in consulting or teaching.
TABLE圖表轉述範例Example: describing a chart
不要只寫「此處有一張圖」。應該寫清楚：圖表想比較什麼、橫軸縱軸代表什麼、哪個關係最重要、讀者應該得到哪個判斷。Do not write only “there is a chart here.” Explain what it compares, what each axis means, which relationship matters, and what judgment the reader should take away.
好轉述會讓 AI 之後能引用圖表，而不是只知道文件曾經有圖。A good description lets AI cite the chart later, not merely know that a chart existed.
05
不要只做摘要，要抽出方法Do not stop at summaries. Extract methods.
摘要回答「內容說了什麼」。方法回答「在什麼情境下，為什麼這樣做」。A summary says what the content said. A method explains when to use it and why it works.
真正可重複使用的萃取結果，應包含情境、問題類型、判斷依據、提問策略、適用邊界與代表性範例。A reusable extraction should include context, problem type, judgment criteria, questioning strategy, boundaries, and representative examples.
SHIFT把摘要改寫成方法Convert summary into method
摘要句：「團隊需要重算市場規模。」方法句：「當團隊同時談 B2B 與 B2C，導致定價與財務估算混在一起時，先要求切出第一階段買單者，再重算市場規模。」Summary: “The team needs to recalculate market size.” Method: “When B2B and B2C are mixed and pricing becomes unclear, first choose the initial paying buyer, then recalculate market size.”
方法必須包含觸發條件、判斷理由與下一步動作。A method needs a trigger condition, reasoning, and next action.
CARD可重用方法卡格式Reusable method-card format
- 方法名稱：讓未來可以搜尋與呼叫。Method name: make it searchable and callable.
- 使用條件：什麼情境可以用、什麼情境不要用。Use condition: where it applies and where it should not.
- 操作步驟：先問什麼、接著看什麼、最後要求什麼交付。Steps: what to ask first, what to inspect next, what delivery to request.
- 代表片段：放入 1 到 3 段真實對話作為 few-shot。Representative clips: include 1 to 3 real dialogue snippets as few-shots.
摘要與方法的差別Summary versus method
外部 Skill 要先檢查，再個人化Inspect external Skills before personalizing
06
Skill 不是提示詞，是一包工作方法A Skill is a working method, not just a prompt
一個好的技能包應該包含使用情境、流程、範例、限制條件、輸出格式與人工檢查點。A good Skill includes context, workflow, examples, constraints, output format, and review gates.
使用外部 Skill 時，先看它是否會讀取敏感資料、執行不明指令，或只是套用通用模板。確認後，再加入自己的專業判斷。Before using an external Skill, check whether it reads sensitive data, runs unclear commands, or only applies generic templates. Then add your own judgment.
SAFE外部 Skill 安全檢查External Skill safety check
- 先看入口文件：它要求 AI 做什麼、何時讀資料、何時呼叫工具。Read the entry file first: what it asks AI to do, when it reads files, and when it calls tools.
- 檢查是否掃描金鑰、讀取環境變數、改寫系統設定、刪除檔案或連到不明服務。Check whether it scans keys, reads environment variables, changes system settings, deletes files, or connects to unclear services.
- 確認輸出物是否有品牌誤導、機密外流或未註明來源的風險。Check risks around brand confusion, confidential leakage, and unsupported sources.
OWN個人化不是換字，而是放入判斷Personalization means adding judgment
顧問型 Skill 不能只寫「請用 MECE 分析」。它要知道你會先判斷經營者心態、確認對方是否能理解陌生問題，再決定要溫和接住、精準追問或要求證據。A consulting Skill cannot only say “use MECE.” It should know that you first assess the leader’s mindset, check whether they can handle unfamiliar problems, then decide whether to acknowledge, probe, or ask for evidence.
好的 Skill 會把你的工作哲學變成使用條件、流程、few-shot、禁止事項與檢查點。A good Skill turns your philosophy into conditions, workflow, few-shots, prohibitions, and review gates.
07
本機工作流讓 AI 不只聊天Local workflows move AI beyond chat
用四個資料夾管理工作：原始資料、處理流程、輸出成果、參考資料。README 則記錄命名規則、處理步驟與輸出格式。Manage work with four folders: Raw Files, Process, Output, and References. README records naming rules, steps, and output formats.
當流程穩定後，可以讓排程定期檢查新檔案，呼叫正確指令，把成果放到固定位置。Once stable, a scheduled task can check new files, call the right instructions, and place outputs in the right folder.
FOLDER四個資料夾各自負責什麼What each folder owns
- 原始資料：逐字稿、PDF、簡報、會議紀錄，保持未加工狀態。Raw Files: transcripts, PDFs, slides, meeting records, kept unprocessed.
- 處理流程：Prompt、Skill、清理規則、中間檔、任務判斷邏輯。Process: prompts, Skills, cleanup rules, intermediate files, task logic.
- 輸出成果：會議紀錄、知識摘要、顧問報告、簡報、教材。Output: meeting records, knowledge summaries, consulting reports, decks, teaching material.
- 參考資料：書籍、範本、過往個案、長期維護的知識庫。References: books, templates, past cases, long-term knowledge bases.
READMEREADME 是交接文件README is the handoff contract
README 要寫給下一個 AI 代理或下一台電腦看：資料夾用途、檔名規則、處理順序、什麼情境呼叫哪個 Skill、輸出要放在哪裡。The README is for the next AI agent or another computer: folder purpose, naming rules, processing order, which Skill to call, and where outputs go.
當 README 寫清楚，AI 工作就不再依賴單次聊天記憶。When the README is clear, AI work no longer depends on one chat history.
資料夾就是工作契約Folders are the work contract
知識系統要定期回收A knowledge system improves through review
錯誤輸出Bad outputs留下原因，不只刪掉結果。Keep the cause, not just the failed result.
好範例Strong examples把可重用片段放回 Skill。Move reusable pieces back into the Skill.
新限制New limits補上不適用情境與禁止事項。Add non-use cases and prohibitions.
版本更新Version update讓規則跟著工作現場長大。Let rules grow with the work.
08
每次使用後，都要補回系統After each use, feed the system
個人化知識系統不是一次完成。真正的進步來自回收錯誤、補充案例、修正規則、更新 Skill。A personal knowledge system is never finished. It improves by feeding back errors, cases, rules, and Skill updates.
當 AI 的輸出不符合專業判斷時，不要只重下指令；回到逐字稿、範例與規則，找出哪一段沒有教清楚。When AI output does not match professional judgment, do not only reprompt. Go back to transcripts, examples, and rules to find what was not taught clearly.
REVIEW每次使用後的回收表After-use review sheet
- 這次輸出哪裡不像你的判斷？是語氣、邏輯、格式，還是缺少某個情境？Where did the output miss your judgment: tone, logic, format, or missing context?
- 哪一段原始資料可以補成更好的 few-shot？Which source clip can become a better few-shot?
- 這次失敗代表要新增規則、限制條件，還是禁止事項？Does this failure require a new rule, condition, or prohibition?
LOOP更新週期怎麼抓How often to update
單次任務後只補錯誤與好範例；每週整理高頻問題；每月再合併成新版 Skill。這樣既不會每次大改，也不會讓知識庫停在舊規則。After one task, add only errors and strong examples. Weekly, review frequent issues. Monthly, merge them into a new Skill version. This avoids constant rewrites while keeping the system current.
版本紀錄建議寫：新增了什麼、修正了什麼、哪個情境不再適用。Version notes should say what was added, what was fixed, and which context no longer applies.
下載與複製工作材料Download and copy the working materials
以下材料可直接下載或複製，適合放進 ChatGPT Project、Claude Project、NotebookLM，或本機 AI 工作資料夾。 These materials can be downloaded or copied into ChatGPT Projects, Claude Projects, NotebookLM, or a local AI workflow folder.
搭配操作影片看，會更容易理解資料怎麼流動Use the walkthrough video to see how files move
建議邊看影片邊對照三個位置：原始資料放哪裡、處理指令放哪裡、最後輸出回到哪個資料夾。外部影片可以作為工具補充；這裡先保留與本方法一致的完整操作影片。Watch it with three locations in mind: where raw files go, where process instructions live, and where outputs return. External videos can supplement tool usage; this page keeps the walkthrough aligned with this method.
文件清理指令Document cleanup instructions
把 PDF、書籍、講義清成可閱讀的 Markdown。Turn PDFs, books, and handouts into readable Markdown.
核心概念萃取指令Core concept extraction instructions
整理章節功能、概念關係、適用邊界與知識卡。Extract chapter functions, concept relationships, boundaries, and knowledge cards.
課堂筆記 PromptClass-notes prompt
把課程、討論與引導過程整理成可再使用的筆記。Turn classes, discussion, and facilitation into reusable notes.
會議紀錄 SkillMeeting record Skill
包含萃取與審查兩個 Skill ZIP，適合建立會議處理流程。Includes extractor and review-gate Skill ZIPs for meeting workflows.
台灣中文文風指令Taiwan Chinese writing prompt
把生成文字調整成較自然的台灣商管與教學語氣。Localize generated text into a natural Taiwan business and teaching voice.
快速開始：先清理，再萃取Quick start: clean first, extract second
B：請先清理這份文件，保留章節、表格、圖表意義與無法判讀處。完成後，再用 V 流程萃取核心概念、概念關係、適用情境、不適用情境與知識卡。B: Clean this document first. Preserve chapters, tables, chart meanings, and uncertain passages. Then run V to extract core concepts, concept relationships, use cases, non-use cases, and knowledge cards.
快速開始：從逐字稿抽出方法Quick start: extract method from a transcript
請不要只做摘要。請從這份逐字稿整理出情境、對象、問題階段、關鍵追問、方法選擇、判斷依據、few-shot 範例、禁止事項與後續可更新的 Skill 規則。Do not only summarize. From this transcript, extract context, audience, problem stage, key questions, method choices, judgment criteria, few-shot examples, prohibitions, and Skill rules that can be updated later.
公開頁面收錄的是可教學、可複製的材料。大型影片以外部連結提供，原始逐字稿不放入頁面主視圖，避免干擾學習與造成不必要的公開風險。 This public page includes teaching-ready and reusable materials. Large videos are provided as external links, and raw transcripts are not shown in the main view to keep the learning path clean and reduce unnecessary public exposure.