AI Context Management Lecture Slide Outline
Created: 2026-06-14 Updated: 2026-06-15 Status: Page-by-page slide outline for Marcus review Audience: 公司內非工程同事 Duration: 2 hours
Course Spine
這堂課的主線是:
Skill -> Context -> Memory
核心訊息:
Skill 讓 AI 知道怎麼做事;Context 讓 AI 知道根據什麼背景做事;Memory 讓這些背景能跨時間留下來。
這份文件是逐頁 slide outline。每一頁包含:
- 這一頁的標題。
- 這一頁的起承轉合,也就是這頁要承接什麼、建立什麼概念、轉向哪裡。
- 這一頁預計的內容與合適配圖。
本課不展開 Harness / Loop。Harness / Loop 最多作為 AI 發展脈絡中的背景位置,不作為本次教學主內容。
Suggested Slide Count
主版規劃為 21 頁,搭配 2 小時課程。這次不再把現場練習獨立拆成 Step 1-5 slides,而是改成「講完一個觀念,就讓同事執行對應 Phase」。
| Section | Slides | Time |
|---|---|---|
| Opening | 1-2 | 5 min |
| From Skill to Context | 3-5 | 15 min |
| Context Management Basics | 6-9 | 15 min |
| Personal Context Layer Phases 1-3 | 10-13 | 45 min |
| Context as Memory | 14-16 | 15 min |
| Company Context Service + MCP Phase 4 | 17-19 | 20 min |
| Wrap | 20-21 | 5 min |
Slide 01: Context Management
起承轉合
這是開場頁。先讓大家知道今天是承接上次 Skill 課程,往下一步進到 Context。這頁只需要建立主軸:AI 不只是要有好 prompt 或 skill,而是要能在正確背景裡工作。
預計內容與配圖
內容:
- Title: Context Management
- Subtitle: 從 Skill 到 Context
- 一句話主軸:讓 AI 不只是會做事,而是能根據正確背景做事
- 日期 / 公司內訓 / Marcus
配圖:
- 簡單二段式路徑圖:
Skill -> Context - 或一張「Agent 面前有一包工作背景資料」的抽象圖
Slide 02: 今天會帶走什麼
起承轉合
開場後先講結果,避免大家覺得這只是概念課。這頁要把課程落在具體成果:理解 Context,建立本機 Context Layer,知道 Memory 和公司 Context Service 是什麼。
預計內容與配圖
內容:
- 今天結束後,你會帶走:
- 一個本機 Context Layer folder
- 一份
USER.md - 一套 PARA 資料夾
- Project / Area 的 MOC 文件
- 一個最小可用的
context-managerskill - 一個可以讓 Agent 讀取、整理、回寫 Context 的起始環境
- 知道後續如何用 MCP 連上 Company Context Service
配圖:
- Checklist visual
- 或「今天完成的 7 個成果」卡片
Slide 03: AI 工作流的五個階段:上次到 Skill
起承轉合
這頁承接上次課程,把 Prompt、Skill、Context、Harness、Loop 放在同一條發展路徑上。先指出上次我們走到 Skill:Skill 讓 prompt 從一次性提問變成可重複流程。接著轉折:Skill 會做事,但它仍然不知道你的工作背景,所以今天要往 Context 前進。
預計內容與配圖
內容:
- 五階段路徑:
Prompt -> Skill -> Context -> Harness -> Loop
- Prompt:一次性的直接提問,結果比較隨機
- Skill:結構化、流程化、可重複使用的 prompt
- 上次課程位置:指針停在 Skill
- Skill 固定的是:
- 任務流程
- 產出格式
- 檢查清單
- 語氣與結構要求
- 常見錯誤
- 但 Skill 不一定知道:
- 我是誰、我的角色是什麼
- 這個專案目前到哪裡
- 公司已經決定過什麼
- 主管偏好什麼格式
- 哪些話不能說
- 哪些範例可以直接沿用
轉折句:
Skill 解決「怎麼做」;Context 解決「根據什麼背景做」。
配圖:
- 五步驟水平進度條,指針停在 Skill
- 動畫第二層:Skill 旁邊浮出「背景缺口」
Slide 04: 今天往下一格:Context
起承轉合
這頁接在五步驟之後,讓指針從 Skill 往 Context 移動。它不是要展開 Harness 或 Loop,而是清楚說明:今天的任務就是把 AI 從「會照流程做事」推進到「知道根據什麼背景做事」。
預計內容與配圖
內容:
Prompt -> Skill -> Context -> Harness -> Loop
^
今天
Context 是 AI 執行任務時需要的:
- 背景
- 環境
- 資料
- 判斷標準
- 可回寫知識
Working definition:
Context 是讓 AI 知道現在在哪裡工作、根據什麼判斷、應該產出什麼、產出後哪些東西要留下來的背景系統。
本次課程只深入到 Context:
- Harness:知道位置,不展開教學
- Loop:知道位置,不展開教學
配圖:
- 同一個五步驟水平進度條,指針從 Skill 移到 Context
- Agent 周圍有 identity、project、knowledge、memory 的背景層
Slide 05: Prompt / Skill / Context
起承轉合
把三者放進同一張表,讓大家有清楚分工。這頁是整堂課的重要概念基準。
預計內容與配圖
內容:
| Layer | Simple Definition | Example |
|---|---|---|
| Prompt | 這次我要你做什麼 | 幫我寫一封 email |
| Skill | 這類任務通常怎麼做 | 先判斷對象,再列重點,再產出中英版本 |
| Context | 做這件事時必須知道什麼 | 客戶背景、公司語氣、之前承諾、不能說的話 |
關鍵句:
好的 AI 工作流,不是只有 prompt,也不是只有 skill,而是 prompt + skill + context。
配圖:
- 三層堆疊圖:Prompt 在最上層、Skill 是流程層、Context 是環境層
Slide 06: 五種常見 Context
起承轉合
定義後要拆解類型。這頁讓同事知道 Context 不是單一文件,而是一組不同用途的背景。
預計內容與配圖
內容:
| Type | Meaning | Example |
|---|---|---|
| Identity Context | 我是誰、我的角色、我怎麼工作 | 我是 Marketing,常做社群與活動 |
| Task Context | 這次任務目標、限制、成功標準 | 週報給主管,3 分鐘內看懂 |
| Knowledge Context | 穩定背景知識 | 產品定位、FAQ、品牌語氣 |
| Working Context | 專案進行中的狀態 | 本週進度、未決問題 |
| Memory Context | 過去萃取出的偏好、決策、教訓 | 主管不喜歡太長版本 |
配圖:
- 五張小卡片
- 或 Context toolbox
Slide 07: Context 不是多貼資料
起承轉合
指出常見誤解。這頁要打破「context window 很長,所以我把資料都貼進去就好」的想法。
預計內容與配圖
內容:
常見錯誤:
- 把一堆文件一次貼給 AI
- 期待長 context window 自動解決一切
- 每次重開對話都重新講背景
- 只保存最終產出,不保存決策原因
更好的做法:
- 用檔案保存 context
- 用固定結構整理 context
- 讓 Agent 先讀 context,再開始做事
- 任務完成後讓系統自動維護、更新可複用的背景
配圖:
- 左邊:聊天框裡堆滿文字
- 右邊:Agent 先讀取本機 Context Layer,再開始工作
Slide 08: Context 要能被保存、取用、回寫
起承轉合
從「不是多貼資料」轉到「要管理」。這頁要建立 Context Management 的三個核心動作:保存、取用、回寫。
預計內容與配圖
內容:
Context Management 的三個動作:
- Save:重要背景要留下來
- Retrieve:做事前要找得回來
- Write Back:任務後要知道什麼值得更新
關鍵句:
Context 不是一次性的輸入,而是一個會被持續維護的工作環境。
配圖:
- 循環圖:Save -> Retrieve -> Use -> Maintain
Slide 09: 今天的實作方式:Agent Instruction + Phases
起承轉合
把概念接到實作。這頁說明今天不是讓大家照投影片手動建立資料夾,而是把一份 Agent-facing instruction 交給自己的 Agent。投影片只負責建立心智,每一段講完就進入對應 Phase。
預計內容與配圖
內容:
今天的現場節奏:
| Phase | 目標 | 產物 |
|---|---|---|
| Phase 1 | 建立 Personal Context | USER.md |
| Phase 2 | 建立 PARA + MOC | folders、Project / Area / Resource MOC |
| Phase 3 | 建立 Context Manager | context-manager/SKILL.md |
| Phase 4 | 連上公司 Context | Company Context MCP status |
關鍵句:
Slide 講概念;Agent Instruction 負責執行。
現場同事要打的指令會很短:
請依照 Agent Instruction,開始 Phase N。
每個 Phase 裡真正要問什麼、建立什麼檔案、如何驗證,都寫在 Agent Instruction 裡。Slide 上只顯示 phase trigger,避免大家照抄一長串細節。
配圖:
- Four-phase rail:Concept -> Prompt Box -> Agent Action -> File Output
- 每個 Phase 頁面都放一個短的 terminal-style phase trigger
Slide 10: Phase 1 - Personal Context Layer
起承轉合
正式介紹個人 Context Layer,並立刻進入 Phase 1。這頁要重新框定:這不是筆記整理,而是 AI 可讀、可更新的工作背景層。Phase 1 的任務是用很少的問題建立基本身份 context,最後寫入 workspace root 的 USER.md。
預計內容與配圖
內容:
Personal Context Layer 是:
我的工作背景 + 我的知識資料夾 + 我的 Agent
它要做到:
- 讓 Agent 知道我是誰
- 讓 Agent 知道我正在做什麼
- 讓 Agent 找得到過去資料
- 讓 Agent 任務後能自動維護、更新可複用背景
Phase 1 prompt box:
請依照 Agent Instruction,開始 Phase 1。
Phase 1 的重點:
- Agent 只會先問三個核心問題:名字、角色與職責、常見 AI 產出
- 接著確認兩個 setup 設定:Context 預設位置、主要使用的 Agent
- 同事只要回答 Agent 的問題,不需要照抄設定細節
- 最後會產出一份 AI 可讀的
USER.md
輸出:
AI-Context/
USER.md
配圖:
- User ->
USER.md-> Agent USER.md像一張「工作身份卡」,旁邊連到 responsibilities、outputs、Agent setup
Slide 11: 兩種 Context 結構:PARA 與 Wiki Graph
起承轉合
處理結構選擇。這頁不是要馬上決定誰比較高級,而是讓同事知道:Context 可以用「歸檔邏輯」整理,也可以用「知識網路」長出關係。今天先用 PARA,因為它最容易開始;未來可以慢慢加入 wiki index / MOC,轉成更像 graph 的結構。
預計內容與配圖
內容:
| Structure | 核心問題 | 適合什麼 | 對 Agent 的價值 |
|---|---|---|---|
| PARA | 這個東西現在應該放哪裡? | 工作文件、專案、責任範圍、參考資料 | 快速定位任務狀態與可用材料 |
| Wiki Graph / MOC | 這些知識彼此怎麼連起來? | 主題知識、跨專案洞察、長期演化的理解 | 讓 Agent 能沿著 index、links、backlinks 找到相關脈絡 |
補充說明:
- PARA 是歸檔邏輯:Projects、Areas、Resources、Archives
- Wiki Graph 是連結邏輯:內部連結、MOC、index、backlinks、graph view
- MOC 可以理解成「主題地圖」:一頁索引,把分散的相關筆記串起來
今天的選擇:
今天用 PARA 起步,因為它簡單、好入門,也符合大家日常整理筆記的方式。
後續可以加入 wiki index / MOC,讓資料夾慢慢長成知識網路。
Research anchors:
- Forte Labs PARA:Projects / Areas / Resources / Archives,重點是依行動性與工作狀態整理
- Obsidian internal links / graph view:透過 notes 與 links 形成 knowledge network
- MOC / evergreen notes / Zettelkasten:強調連結、主題索引與跨時間演化
配圖:
- 左半:PARA folder shelf,四個抽屜
- 右半:Wiki Graph,中心是
_Index.md/ MOC,連到多個 note nodes - 中間加一句:
PARA helps you place things. Graph helps you connect things.
Slide 12: Phase 2 - PARA 書櫃與 MOC 地圖
起承轉合
前一頁講兩種結構,這頁把今天要採用的 PARA 講清楚,並進入 Phase 2。重點不是建立漂亮資料夾,而是讓 Agent 問出目前有哪些 Projects / Areas,並幫每個 Project / Area 建立一份 MOC 文件,讓資料夾從第一天就有入口。
預計內容與配圖
內容:
PARA:
| Folder | Purpose | Example |
|---|---|---|
00_Inbox | 還沒整理的輸入 | 會議片段、文章、臨時想法 |
10_Projects | 有明確目標與結束點的工作 | Q3 campaign、活動頁、研究報告 |
20_Areas | 長期負責的職能或領域 | Marketing、HR、客服、內容 |
30_Resources | 反覆引用的知識、範例、資料 | 品牌語氣、FAQ、競品資料 |
40_Templates | 固定格式、prompt、模板、skill 草稿 | 週報模板、會議摘要格式 |
90_Archive | 已完成、過期、暫停內容 | 舊專案、已整理原始資料 |
Phase 2 prompt box:
請依照 Agent Instruction,開始 Phase 2。
預期結構:
AI-Context/
00_Inbox/
10_Projects/
_Projects.md
<Project Name>.md
20_Areas/
_Areas.md
<Area Name>.md
30_Resources/
_Resources.md
40_Templates/
90_Archive/
.agent/
skills/
context-manager/
USER.md
CONTEXT_LOG.md
MOC 的角色:
- 不是要同事手動設計一套 wiki
- Agent 會依照 Instruction 詢問 Projects / Areas,並建立對應入口
- 對同事來說,MOC 就是讓 AI 找得到專案與責任脈絡的地圖
配圖:
- 一個書櫃視覺:
- 書櫃層板是 PARA folders
- 每層最前面有一張 index card:
_Projects.md、_Areas.md、_Resources.md - Project / Area 文件像可抽出的卡片
Slide 13: Phase 3 - Context Manager 是圖書館管理員
起承轉合
前面已經有書櫃與地圖,但還缺一個日常操作介面。這頁用「圖書館管理員」比喻 context-manager:它不是另一份知識,而是幫你收資料、整理資料、匯入資料、取用資料的 skill。
預計內容與配圖
內容:
如果 Context Layer 是書櫃,context-manager 就是圖書館管理員。
四個基本能力:
| Ability | 使用時機 | 會做什麼 |
|---|---|---|
| Inbox | 先收進來 | 把原始內容安全放進 00_Inbox |
| Organize Inbox | 整理收件匣 | 逐筆讀 Inbox,完整 ingest,更新 Project / Area / Resource |
| Ingest | 匯入內容 | 把文章、會議、文件整理成未來可用 context |
| Get Context | 取用背景 | 在產出前先找相關 context,再開始工作 |
Phase 3 prompt box:
請依照 Agent Instruction,開始 Phase 3。
成功標準:
AI-Context/.agent/skills/context-manager/SKILL.md存在- skill 知道 root 的
USER.md - skill 知道如何讀 PARA / MOC
- skill 的 Organize Inbox 不是只給建議,而是整理前先列 plan,確認後執行 ingest 與更新
配圖:
- 左邊是書櫃,右邊是一個 librarian icon
- 四個操作按鈕:Inbox / Organize / Ingest / Get Context
Slide 14: 從 Context 到 Memory
起承轉合
進入 Memory 段。這頁把剛才建立的本機 Context Layer 重新命名為最小可用的個人 AI Memory。
預計內容與配圖
內容:
關鍵句:
Context 是這次任務需要知道的背景;Memory 是這些背景跨時間留下來,並在未來能被重新取用、更新、修正與忘記。
對照:
- Context:這次週報要寫給主管
- Memory:主管偏好先看結果、再看原因、最後看下一步
配圖:
- 左邊:Context 像一個文件夾,裡面放「這次任務」需要的資料
- 右邊:Memory 是有時間維度與知識維度的網狀結構
- 時間軸:過去任務、會議、決策
- 知識節點:偏好、流程、範例、修正
- 箭頭:新的任務會取回舊背景,也會回寫新理解
Slide 15: Memory 不是 Chat History
起承轉合
先排除常見誤解。這頁要讓同事知道 Memory 不是把聊天記錄全部留下,而是萃取未來有用的背景。
預計內容與配圖
內容:
| Chat History | Memory |
|---|---|
| 原始對話紀錄 | 經整理後的可用背景 |
| 很多、雜、容易過期 | 較少、重要、可被維護 |
| 可能只是流水帳 | 能幫助未來決策或產出 |
| 不一定正確取回 | 有結構、來源、更新方式 |
一句話:
Chat history 是過去說過的東西;Memory 是從過去萃取出未來還有用的東西。
配圖:
- Raw transcript vs distilled memory
- 左邊是一長串 chat bubbles,右邊是整理後的 preference / decision / workflow cards
Slide 16: Memory Lifecycle
起承轉合
把 Memory 從名詞帶回操作。這頁要連回剛才建立的 Inbox / Organize / Get Context,而不是再增加新的術語分類。
預計內容與配圖
內容:
Capture -> Organize -> Retrieve -> Use -> Maintain
| Stage | Local Practice |
|---|---|
| Capture | 放進 00_Inbox |
| Organize | Organize Inbox,更新 PARA / MOC |
| Retrieve | Get Context |
| Use | 用 context 完成任務 |
| Maintain | 修正、更新、封存過期內容 |
關鍵句:
Memory 的價值不是記很多,而是能在下一次任務中拿出正確的背景。
配圖:
- Circular lifecycle
- 最後一段 Maintain 回到
USER.md、Project MOC、Area MOC、Resources
Slide 17: Company Context Service:組織級 Context Layer
起承轉合
從個人記憶轉到組織記憶。轉折點是:前面我們做了個人 Context Layer,但如果每個人都要自己整理公司知識、整理同一批會議紀錄、重建同一組決策背景,就很浪費,也容易不一致。所以公司需要一個集中、可信、可治理的 Context Layer。
預計內容與配圖
內容:
個人:我的 Markdown / USER.md / PARA / MOC -> 我的 Agent
公司:公司確認過的知識 / 來源 / 層級 / 權限 -> 所有 Agent
一句話:
個人 Context Layer 解決「我的 AI 要知道我怎麼工作」;Company Context Service 解決「公司的 AI 要根據同一套可信知識工作」。
Context Service 做的事:
- 把 approved KB、HR source、Linear / Project working memory、meeting rollups 編譯成 snapshot
- Runtime / MCP 只讀 snapshot,不直接亂讀原始 Notion、Linear、Slack 或 raw files
- Agent 拿到的是 governed Context Pack,而不是自由瀏覽公司所有資料
配圖:
- 如果 instructor 之後提供 Visualizer screenshot,就使用 screenshot
- 如果沒有 screenshot,就用 diagram fallback:
- 中心:SSOT source nodes
- 外圈:L1 / L2 / L3 CTX nodes
- 右側或下方:Query panel -> Context Pack
Slide 18: Context Service 的幾個詞
起承轉合
在 MCP 前先定義名詞。這頁不講工程細節,只講同事看 demo 時需要理解的語言,並用一張圖說明資料如何從 source 變成 Agent 可用的 Context Pack。
預計內容與配圖
內容:
| Term | 白話說法 |
|---|---|
| SSOT | 原始可信來源,知識最初被確認的地方 |
| L1 Evidence | 直接證據,例如會議紀錄、source excerpt、事件證據 |
| L2 Working Memory | 近期狀態、決策、open questions |
| L3 Living Knowledge | 較穩定的知識、流程、原則 |
| Snapshot | 經過編譯、清理、權限治理後的可查資料包 |
| Context Pack | Agent 回答前拿到的一包可信背景 |
| MCP | Agent 查公司記憶的標準入口 |
關鍵句:
L1 是證據,L2 是近況,L3 是穩定知識;它們不是高低,而是用途不同。
配圖描述:
- 一張由左到右的 pipeline 圖:
SSOT Sources
-> Compiler + Sanitizer
-> Snapshot
-> Governed Query Tools
-> Context Pack
-> Agent Answer
- 在 Snapshot 區塊下方加一個 graph inset:
- 金色節點:SSOT documents
- 紅橘節點:L1 evidence
- 紫色節點:L2 working memory
- 藍色節點:L3 living knowledge
- 金色邊:source binding / evidence relationship
- 在 Agent Answer 旁邊標註:
有來源、有層級、有權限邊界
Slide 19: Phase 4 - 替你的小腦連上公司的大腦
起承轉合
用非工程語言介紹 MCP,並進入 Phase 4。這頁的重點是:同事不需要記兩套操作方式。最後帶走的是一個 context-manager skill,它同時能管理個人的 Context Layer,也能在公司問題上透過 MCP 查 Company Context Service。
預計內容與配圖
內容:
MCP 對使用者來說:
讓 Agent 查公司記憶的標準插座
概念:
我的問題
-> context-manager skill
-> 判斷是個人 context 還是公司 context
-> Personal Context Layer 或 Company Context MCP
-> Agent 回答
Phase 4 prompt box:
請依照 Agent Instruction,開始 Phase 4。
Phase 4 原則:
- 本機 Context Layer 成功後才嘗試
- MCP setup 細節已寫在 Agent Instruction 裡
- 需要使用者確認
- 需要 OAuth / SSO 時,由同事自己完成瀏覽器登入
- 成功、跳過、失敗都要明確回報狀態
成功標準:
- 同事只需要記得一個 skill:
context-manager - 個人問題:用本機
USER.md/ PARA / MOC - 公司問題:用 Company Context MCP
- 公司 context 的答案要保留來源 / Context Pack / 權限邊界
配圖:
- 小腦 / 大腦橋接圖:
Personal Context Layer
|
context-manager
|
Company Context MCP -> Company Context Service -> Context Pack
Slide 20: 今天完成了什麼
起承轉合
進入收尾。這頁要依照前面的實際內容重新盤點成果:不是單純建立資料夾,而是把同事從 Skill 帶到 Context,再讓他知道 Context 如何變成 Memory,最後看見個人與公司 Context Layer 如何接起來。
預計內容與配圖
內容:
今天完成:
- 理解 Prompt / Skill / Context 的差別
- 知道 Context 不是多貼資料,而是要保存、取用、維護
- 建立 Phase 1:
USER.md - 建立 Phase 2:PARA folders + Project / Area / Resource MOC
- 建立 Phase 3:
context-managerskill - 理解 Context 如何變成 Memory
- 看懂 Company Context Service 的基本概念:SSOT、L1、L2、L3、Snapshot、Context Pack
- 建立 Phase 4 的心智:用 MCP 讓同一個 skill 查公司 Context
配圖:
- Completion checklist
- 或一條路徑圖:Skill -> Personal Context Layer -> Memory -> Company Context Service
Slide 21: 下課後可以直接用的 Prompts
起承轉合
給同事立刻可用的下一步。這頁讓 Context Layer 不停在課堂 demo,而能回到日常工作。
預計內容與配圖
內容:
Personal context prompts:
請幫我製作一份 Forest 競品分析 research,做成 Markdown file 加入 Inbox。
請幫我整理 Inbox。請逐一讀取 Inbox 內容,完整 ingest,並把可以歸檔的內容整理進 Project、Area 或 Resource,最後更新 CONTEXT_LOG。
請先取得和這個任務相關的 Context,再開始幫我寫。
請根據我的 USER.md、Project MOC 和 Area MOC,幫我產出這週的工作更新。
請把這份會議紀錄 ingest 到 Context Layer,整理成決策、待辦、open questions,以及應該更新到哪個 Project / Area MOC。
Company context prompts:
請用公司 Context Service 查詢:2026 下半年重要目標是什麼?
請用公司 Context Service 查詢:目前 Context Layer 專案推進到哪裡?
請先查公司 Context,再幫我整理這個主題目前公司已經確認過的事實、仍不確定的問題,以及我下一步應該問誰。
備註:
- Company Context prompts 只在
company_mcp_connected時使用 - 如果是
company_mcp_skipped或company_mcp_failed,Agent 應明確回報公司 Context 目前不可用,不要猜答案
配圖:
- Prompt cards grouped by personal context / company context
Research References
- Forte Labs: PARA Method
- Obsidian Help: Internal links
- Obsidian Help: Graph view
- Linking Your Thinking / Nick Milo: Maps of Content
- Andy Matuschak: Evergreen notes
- Zettelkasten.de: Introduction to the Zettelkasten Method
- Local reference:
vault/CLAUDE.md - Local reference:
services/company-context/README.md - Local reference:
docs/architecture/company-context-graph-visualizer.md
Claude Design Implementation Brief
This section exists for a separate design/build agent, such as Claude Design, that does not have access to the previous conversation or this workspace's broader context. Treat this file as the source of truth for the slide deck design and content.
The companion Agent Instruction file is the source of truth for live exercise execution details. Claude Design does not need to reproduce or redesign that instruction file. For slide design, use the summary in this brief and the page-by-page outline below.
What This Deck Is
This is a 2-hour internal company lecture deck about Context Management for non-engineering colleagues. The course is the second step after a previous lesson about Skills. Colleagues already have some practical exposure to Claude Code, Codex, Cursor, or similar codebase agents, and may be able to write simple skills. They are not expected to be engineers.
The deck should help them move from:
I can ask AI to do a task
-> I can give AI a reusable skill
-> I can create and maintain a local Context Layer so AI knows my working background
The teaching spine is:
Skill -> Context -> Memory
The core message is:
Skill 讓 AI 知道怎麼做事;Context 讓 AI 知道根據什麼背景做事;Memory 讓這些背景能跨時間留下來。
What The Lecture Must Achieve
By the end of the class, colleagues should:
- Understand the difference between Prompt, Skill, Context, and Memory.
- Understand why a skill alone is not enough if the AI lacks working background.
- Create a small local Context Layer on their own computer.
- Understand why the local Context Layer starts with
USER.md, PARA folders, and Project / Area / Resource MOC files. - Create or install a minimal
context-managerskill that can capture Inbox items, organize Inbox, ingest content, and get context before doing a task. - Understand that Context can become Memory when it is saved, retrieved, used, maintained, corrected, and forgotten over time.
- Understand, at a conceptual level, why the company also has a larger Context Service / Memory layer.
- Understand that Company Context Service can be accessed by the same
context-managerskill through MCP when configured.
The class is practical. The deck should not feel like a theoretical AI lecture. It should feel like: "Here is the next operating habit you can start using today."
Companion Document: Agent Instruction
There is a separate file:
vault/00_Inbox/AI Lecture/Agent Instruction.md
That file is not the slide deck. It is an Agent-facing bootstrap instruction that colleagues will give to their own Agent during the live exercise.
The deck should refer to the Agent Instruction as the execution layer:
Slide 講概念;Agent Instruction 負責執行。
Important: do not copy long operational instructions from the Agent Instruction into the slides. The executable prompt box on each Phase slide should contain only the short phase trigger. The slide may also show the phase purpose, expected output, and why it matters. The detailed behavior, questions, file templates, safety rules, and verification steps live in the Agent Instruction.
The live operation has four phases:
| Phase | Human-facing trigger shown on slide | What the Agent Instruction handles |
|---|---|---|
| Phase 1 | 請依照 Agent Instruction,開始 Phase 1。 | Ask name, role / responsibilities, common AI outputs, context path, and Agent runtime, then create USER.md. |
| Phase 2 | 請依照 Agent Instruction,開始 Phase 2。 | Create PARA folders and Project / Area / Resource MOC files. |
| Phase 3 | 請依照 Agent Instruction,開始 Phase 3。 | Create the portable context-manager skill. |
| Phase 4 | 請依照 Agent Instruction,開始 Phase 4。 | Connect Company Context Service MCP using the companion Agent Instruction, handle OAuth / SSO, and validate. |
Intended In-Class Flow
The class alternates between concept and action:
Explain a concept
-> show one short Phase trigger
-> colleagues ask their own Agent to run that Phase
-> return to the deck for the next concept
The deck should therefore make each Phase page very scannable. A colleague should know:
- What this phase is for.
- What they should type.
- What output they should expect.
- Why it matters.
But the deck should not look like a command manual.
Audience And Tone
Audience:
- Company colleagues, mostly non-engineers.
- Comfortable enough with AI agents to run simple agent instructions.
- Not expected to understand MCP, SQLite, LLM retrieval, or code architecture.
Tone:
- Traditional Chinese.
- Clear, concrete, practical.
- Avoid engineering-heavy jargon unless the slide immediately translates it into plain language.
- Use analogies carefully:
- Context Layer as a workspace / bookshelf / map.
context-manageras a librarian.- Memory as a time-aware knowledge network.
- MCP as the standard socket for connecting an Agent to company memory.
Visual Direction
Design the deck as a polished internal operating-system style presentation, not a marketing deck.
Suggested visual language:
- Clean, structured, information-rich slides.
- Use diagrams, layer maps, flow charts, folder structures, and simple visual metaphors.
- Use terminal-style prompt boxes only for short phase triggers.
- Use a consistent progression:
- Skill as process
- Context as environment
- Local Context Layer as personal workspace
- Memory as time + knowledge network
- Company Context Service as organization-level memory
- Avoid pages that are only dense bullet lists. Convert dense parts into tables, diagrams, or annotated structures.
- Keep code / command text minimal and readable.
- Use visual contrast between:
- Personal local context
- Company-level context
- Human action
- Agent execution
Suggested recurring motifs:
- A horizontal path:
Prompt -> Skill -> Context -> Harness -> Loop, with the lecture focused only on Context. - A folder shelf for PARA.
- A MOC / index card as the map inside the shelf.
- A small "librarian" or operator icon for
context-manager. - A context-as-folder vs memory-as-network visual.
- A company graph / observatory visual for Company Context Service.
Visual System Contract
Use a consistent visual system across all 21 slides.
Format:
- 16:9 presentation format.
- Projection-first readability.
- Prefer clean full-slide layouts over decorative cards.
- Use tables only when comparison is the point; otherwise use diagrams or annotated structures.
Suggested slide templates:
| Template | Use For | Structure |
|---|---|---|
| Concept Frame | Definitions and key shifts | Large title, one sentence thesis, 2-3 supporting points, simple visual |
| Comparison | Prompt / Skill / Context, PARA / Wiki Graph, Chat History / Memory | Two-column or three-column table with one strong takeaway |
| Phase Slide | Phase 1-4 live operation | Phase label, why it matters, short trigger, expected output, state cues |
| System Diagram | Context lifecycle, Company Context Service, MCP | Left-to-right or center-out diagram with labeled nodes |
Color roles:
- Personal local context: calm green or teal.
- Company context: blue or indigo.
- Human action: warm amber.
- Agent execution: cyan.
- Memory / time dimension: violet.
- Evidence / SSOT: gold.
- Warning / blocked / skipped state: muted red or orange.
Do not rely on color alone. Pair colors with labels, icons, or line styles.
Typography:
- Use large, presentation-safe type.
- Keep titles short and strong.
- Body text should be scannable from a meeting-room projector.
- Terminal trigger boxes should be visually distinct but not tiny.
- Avoid dense paragraphs on slides; convert to diagrams or short labels.
Diagram grammar:
- Use arrows for process.
- Use stacked layers for Prompt / Skill / Context and personal vs company context.
- Use folder shelf visuals for PARA.
- Use index-card / map visuals for MOC.
- Use node-link network visuals only for Memory or Company Context Service.
- Use prompt-card visuals only for the short Phase triggers and post-class starter prompts.
Phase Slide State Pattern
Each Phase page should support live classroom execution. Do not create separate waiting or troubleshooting slides; include small state cues on the Phase slide itself.
For Phase 1-4 slides, include:
| State | What To Show |
|---|---|
| Start | Short trigger phrase in a terminal-style box |
| Working | Small cue: Agent is asking / creating / verifying |
| Done | Expected output checklist |
| Blocked | One short fallback cue, such as Ask instructor / Use manual fallback / Skip MCP for now |
| Resume | A visual cue that brings everyone back to the deck after the Agent finishes |
Examples:
- Phase 1 done:
USER.md created - Phase 2 done:
PARA folders + MOC created - Phase 3 done:
context-manager skill created - Phase 4 done:
MCP connected / skipped / failed with reason
The goal is to help the instructor manage a room where some people finish early, some wait for Agent questions, and some hit setup issues.
Projection And Accessibility Constraints
This deck will be used in a live 2-hour session. Design for room readability.
- One primary concept per slide. A slide may contain 2-3 supporting beats when those beats build the same concept; reveal them progressively instead of showing all content at once.
- Maximum 3-5 visible bullets on most slides.
- Avoid tiny tables; if a table has many words, turn it into cards or a diagram.
- Phase trigger boxes must be large enough to read and copy.
- Use high contrast for all text.
- Do not encode meaning by color alone.
- Label diagram nodes directly.
- Keep icons simple and familiar.
- Avoid decorative visuals that compete with the teaching point.
- If a slide has too much source content, prioritize the slide's main idea and move supporting detail out of the visible slide design.
Progressive Reveal And Animation Contract
Because this will be an HTML slide deck, use simple progressive reveal where a slide has one large concept but multiple teaching beats. The goal is not decorative motion; it is pacing.
Interaction:
- Support keyboard/click progression through reveal steps.
- Use simple fades, highlights, slide-ins, or diagram build steps.
- Avoid complex motion that distracts from the lecture.
- Do not require the instructor to precisely time animations.
- Each reveal state should still look like a complete, readable slide.
- If a slide has no natural sequence, keep it static.
Rule of thumb:
One slide = one main idea.
Multiple beats inside that idea = progressive reveal.
Recommended progressive reveal slides:
| Slide | Reveal Beats |
|---|---|
| Slide 03 | Five-stage path -> Skill fixed process -> background gaps -> "Skill 解決怎麼做;Context 解決根據什麼背景做" |
| Slide 04 | Move pointer from Skill to Context -> define Context -> show Harness / Loop as out-of-scope context |
| Slide 05 | Prompt row -> Skill row -> Context row -> combined takeaway |
| Slide 06 | Reveal context types as grouped cards, not all at once |
| Slide 07 | Common mistake side -> better practice side |
| Slide 08 | Save -> Retrieve -> Write Back -> maintained environment takeaway |
| Slide 09 | Phase rail -> "Slide 講概念;Agent Instruction 負責執行" -> short trigger pattern |
| Slide 10 | Personal Context Layer concept -> Phase 1 trigger -> expected output USER.md |
| Slide 11 | PARA side -> Wiki Graph / MOC side -> today's choice: PARA first, graph later |
| Slide 12 | PARA shelf -> Phase 2 trigger -> expected folder / MOC output |
| Slide 13 | Bookshelf exists -> librarian/context-manager appears -> four abilities -> Phase 3 trigger |
| Slide 14 | Context as folder -> Memory as time + knowledge network -> definition |
| Slide 15 | Chat History side -> Memory side -> distilled-memory takeaway |
| Slide 16 | Capture -> Organize -> Retrieve -> Use -> Maintain, then final value statement |
| Slide 17 | Personal Context Layer -> Company Context Service -> why centralization matters |
| Slide 18 | Terms introduced only as needed -> pipeline builds left-to-right -> governance/source boundary highlight |
| Slide 19 | MCP as socket -> routing personal vs company question -> Phase 4 trigger and status outcomes |
| Slide 20 | Completion checklist reveals by learning arc: Skill/Context -> local layer -> Memory -> Company Context |
| Slide 21 | Personal prompts group -> Company Context prompts group -> availability note |
Slides 01 and 02 can be mostly static unless the designer wants a subtle opening reveal.
Hard Scope Boundaries
Do not expand these into full teaching sections:
- Harness
- Loop
- Four memory type taxonomy
- Memory hygiene checklist
- Standalone L1/L2/L3 deep dive
- Standalone Context Pack deep dive
- A 7-day practice plan
- Academy notification or announcement content
Harness and Loop may appear only in the five-stage AI workflow map to show where this lecture sits. The lecture itself is about Context and the bridge into Memory.
Company Context Service Context
The company has an internal Company Context Service. The instructor will demo a visualizer during the lecture.
Plain-language explanation for slides:
個人 Context Layer 解決「我的 AI 要知道我怎麼工作」。
Company Context Service 解決「公司的 AI 要根據同一套可信知識工作」。
Architecture summary for design purposes:
Approved company sources
-> compiler + sanitizer
-> read-only snapshot
-> governed query tools
-> governed Context Pack
-> Agent answer
Designer note: the underlying tool names are ctx_query, ctx_get, ctx_resolve, and ctx_catalog, but do not put these raw tool names prominently on audience-facing slides unless the slide needs a small technical annotation. For non-engineering colleagues, label this layer as "governed query tools" or "公司 Context 查詢工具".
Terms to define briefly:
- SSOT: 原始可信來源,知識最初被確認的地方。
- L1 Evidence: 直接證據,例如會議紀錄、source excerpt、事件證據。
- L2 Working Memory: 近期狀態、決策、open questions。
- L3 Living Knowledge: 較穩定的知識、流程、原則。
- Snapshot: 經過編譯、清理、權限治理後的可查資料包。
- Context Pack: Agent 回答前拿到的一包可信背景。
- MCP: Agent 查公司記憶的標準入口。
Visualizer design context:
- Existing visualizer is an owner-facing graph view of the compiled Company Context snapshot.
- It uses a "deep-space observatory" style:
- dark field
- gold SSOT nodes
- L1 ember / L2 violet / L3 blue context nodes
- query panel showing how a question becomes a Context Pack
- No screenshot asset is provided in this outline. If the instructor later provides a visualizer screenshot, use it on the Company Context Service slide. Otherwise, use the diagram fallback below.
SSOT sources at center
-> L1 / L2 / L3 rings around them
-> query panel on one side
-> Context Pack flowing to Agent answer
Do not imply that all employees can browse raw Slack, Notion, Linear, or source files. The point is governed access through snapshots and MCP.
What The Deck Should Not Do
- Do not make colleagues copy long prompts from slides.
- Do not make the live exercise look like manual folder setup.
- Do not treat Obsidian as the core concept. Obsidian is only a convenient viewer for Markdown files.
- Do not teach MCP as an engineering protocol. Teach it as the standard connection point between an Agent and company context.
- Do not present Company Context Service as a replacement for personal context. It complements personal context.
- Do not overpromise that memory is automatic or always correct. Emphasize maintenance, correction, and source grounding.
- Do not invent company facts, goals, credentials, or setup commands. MCP setup details live in the companion Agent Instruction; slides should only summarize the concept and show the short Phase 4 trigger.
Expected Deck Output
Create a 21-slide deck following the page-by-page outline below.
Deliverable contract:
- Output should be an HTML slide deck, not a lecture script, handout, generic website, PPTX, or rewrite of the Agent Instruction.
- Preferred file name:
AI Context Management Lecture.html. - If operating inside this workspace, save the HTML deck in
vault/00_Inbox/AI Lecture/. - The HTML should be directly presentable in a browser, with keyboard/click navigation and progressive reveal support.
- Keep the deck self-contained where practical. If separate assets are necessary, place them in a clearly named sibling assets folder under
vault/00_Inbox/AI Lecture/. - PDF export is optional and secondary; the HTML deck is the primary deliverable.
- Use 16:9 presentation format.
- Main slide language should be Traditional Chinese.
- Keep slide text concise. Use the outline's "起承轉合" sections to understand the intended conceptual move, not as text to paste onto slides.
- Do not add speaker notes unless explicitly requested later.
- Do not add a separate appendix unless explicitly requested later.
- Do not create extra live-exercise slides beyond the four Phase pages already specified.
- Keep text and diagrams editable in HTML/CSS/SVG rather than flattening them into screenshots.
Each slide should preserve:
- The title.
- The intended conceptual move.
- The main content.
- The recommended visual idea, adapted into actual slide design.
Slides should be presentation-ready but not overloaded. When the outline contains more text than one slide can comfortably show, prioritize the main teaching point and turn the rest into visual structure or speaker-facing support.
Key Review Criteria
Before finalizing the deck, check:
- Can a designer who did not see the previous conversation understand what this lecture is trying to accomplish?
- Is the practical live flow clear: concept -> short phase trigger -> Agent Instruction executes?
- Are long operational details kept out of the slides?
- Does the deck clearly bridge Skill -> Context -> Memory?
- Does it avoid teaching Harness / Loop beyond positioning?
- Does it explain PARA vs Wiki Graph without making that section too academic?
- Does it make Company Context Service understandable without exposing implementation noise?
- Does every Phase page show only a short trigger and expected outcome?