Sofia Yan

The AI Coworker Methodology · AI 同事方法論

把 AI 當同事，不是當工具 Bring AI onto your team — not as a tool

Sofia Yan（嚴世紀）在 Numbers Protocol 七年用出來的「AI 同事方法論」。 6 個有名字的 AI 同事、5 步導入 playbook、3 個常見失敗模式。 這份方法論為什麼 work？因為它把 AI 當新進團隊夥伴處理，不是當另一個 SaaS。 Sofia Yan's in-production methodology from Numbers Protocol. Six named AI coworkers, a 5-step adoption playbook, three common failure modes. The methodology works because it treats AI as a new team hire, not as another SaaS subscription.

6 named AI coworkers 5-step playbook 3 failure modes Numbers in-production field notes

起點：mindset，不是 model The starting point: mindset, not model

多數企業 AI 導入失敗的原因不是模型不夠強，是 mindset 不對。 「工具」mindset 讓你追求完美、責怪不完美、最後退回 Excel。 「同事」mindset 讓你定義 scope、補不完美、最後升任成 AI 主管。 Most enterprise AI adoptions fail not because the model is weak — but because the mindset is wrong. The "tool" mindset makes you chase perfection, blame imperfection, and retreat to Excel. The "teammate" mindset makes you define scope, fill in gaps, and end up as an AI manager.

這是「8% / 92%」框架的應用版，人類靈魂 8% × AI 產出 92%。 看 Pillar 文完整框架 → This is the operational layer of the "8% / 92%" framework — 8% human soul × 92% AI output. Read the Pillar long-form →

我家 6 位 AI 同事 The 6 AI coworkers

每個 AI 同事都有名字、scope、「不准做的事」。 這份規矩就是把「AI 同事」變成可運作團隊的關鍵。 Each AI coworker has a name, a defined scope, and an explicit "not allowed" boundary. This discipline turns "AI agents" into a functioning team.

Amy

行銷草稿 Marketing Draft

Scope: 長文 blog、社群貼文、電子報草稿、campaign 文案 Long-form blogs, social posts, newsletter drafts, campaign copy

Not allowed: 最終 hashtag 或 Numbers 的數字宣稱（一律真人 verify） Final hashtags or claims about Numbers' metrics (always human-verified)

Jordan

BD 外聯 BD Outreach

Scope: Cold email 草稿、partnership intro 信、follow-up 序列 Cold email drafts, partnership intro emails, follow-up sequences

Not allowed: 實際寄送（hook 攔截，Sofia 逐封 approve） Sending final emails (hook-gated, Sofia approves each)

Clara

社群經營 Community

Scope: Discord / Telegram / Slack 日常營運、FAQ 回應、輕度 moderation Discord / Telegram / Slack daily ops, FAQ replies, light moderation

Not allowed: 危機溝通（偵測到負面 tone 自動 escalate 給真人） Crisis communication (escalates to humans on tone red flags)

小鳳

資料整理 Data Cleaning

Scope: 試算表正規化、去重、lead 增補、CRM 資料整理 Spreadsheet normalization, deduplication, lead enrichment, CRM hygiene

Not allowed: 寫入正式 DB（設計上唯讀） Writes to production DB (read-only by design)

阿張

程式 review Code Review

Scope: PR review 評論、基礎重構建議、依賴 audit PR review comments, basic refactor suggestions, dependency audit

Not allowed: 直接 merge 到 main（永遠需要真人 approver） Direct merge to main (always requires human approver)

老科

文件校稿 Doc Proofreading

Scope: 文法、tone 一致性、中英互譯、style guide 守則 Grammar, tone consistency, translation EN ↔ zh-TW, style guide enforcement

Not allowed: 實質內容變更（只校稿，不重寫） Substantive content changes (only proofs, doesn't rewrite)

5 步導入 playbook The 5-step adoption playbook

1

幫你的 AI 取名字 Name your AI

取一個真人名字（Amy，不是「AI 工具 1 號」）。命名這個動作會逼你的大腦把它當同事，不是工具。名字也強迫你定義「一個 agent 一個 scope」，你不會雇用一個人類處理所有事。 Pick a human name (Amy, not 'AI Tool 1'). The act of naming triggers your brain to treat it as a teammate, not a utility. Names also force you to define a single scope per agent — you wouldn't hire one human to do everything.

2

定義 scope，不是 capability Define scope, not capability

不要說「Amy 會做行銷」。說「Amy 寫 2000 字內的 blog 草稿；她不寫客戶 email」。Scope 是新進真人 day-one 需要的；AI 同事一樣。 Don't say 'Amy can do marketing.' Say 'Amy drafts blog posts up to 2000 words; she does not write client emails.' Scope is what new humans need on day one. AI coworkers need the same.

3

判斷之前先連用一週 Use it daily for one week before judging

多數 AI 導入死在第一週，因為真人期待 day-one perfection。新人有一個月適應期，AI 也該有。注意它做錯的地方，那是 onboarding signal，不是 failure signal。 Most AI rollouts fail at week one because humans expect day-one perfection. New hires get a month. Give AI the same — and pay attention to what it gets wrong (that's onboarding signal, not failure signal).

4

建 audit trail Build the audit trail

每個 AI 同事的動作要有 log：做了什麼、何時、為誰、用什麼 input。這是 TAEA 的「A」(Auditable)。沒它你沒辦法讓新人接 AI 的工作，也沒辦法修錯。 Every AI coworker action needs a log: what it did, when, for whom, with what input. This is the 'A' in TAEA — Auditable. Without it you can't onboard new humans onto the AI's work, and you can't fix mistakes.

5

讓人類同事成為 AI 同事的管理者 Promote humans to 'AI coworker managers'

你的真人同事不會被取代。他們開始管理 AI 同事：設計 scope、review 產出、訓練新行為。工作重心會從手動完成每件事，轉向管理 AI 協作能穩定產出的條件。 Your humans don't get replaced. They manage AI teammates by designing scope, reviewing output, and training new behaviors. The work shifts from doing every task manually to managing the conditions for good AI-assisted work.

3 個常見失敗模式 3 common failure modes

失敗模式 1：把 AI 當「強化版 autocomplete」 Failure 1: Treating AI as 'autocomplete on steroids'

症狀：每次產出都從零開始，沒有記憶、沒有人格。解法：先裝 persistent memory（Claude Dreaming、GPT memories、自家 RAG）再 scale。 Symptom: every output starts from scratch, no memory, no character. Fix: install persistent memory (Claude's Dreaming, GPT memories, custom RAG) before scaling.

失敗模式 2：一個「AI Agent」做所有事 Failure 2: One 'AI Agent' that does everything

症狀：prompt 越來越長，產出品質下滑。解法：拆成 3～5 個各有名字、單一職責的 agent，類似真人版 microservices。 Symptom: prompts get longer and longer, output quality drops. Fix: split into 3-5 named agents with single-responsibility, like microservices for humans.

失敗模式 3：沒有 audit trail、沒有 onboarding 路徑 Failure 3: No audit trail, no onboarding pathway

症狀：只有一個人懂 AI 怎麼運作；他離職全部崩潰。解法：每個 AI 同事都要有書面化 scope + prompt template + 產出樣本 log。 Symptom: only one person knows how the AI works; if they leave, everything breaks. Fix: every AI coworker has a documented scope, prompt template, and output sample log.

TAEA、Z App 與 Omni 怎麼接到日常工作 How TAEA, Z App, and Omni connect to daily work

AI 同事方法論不只是在幫 agent 取名字；真正關鍵的是把每一次交辦、輸出、review 和修正留下脈絡。Numbers Protocol 的 TAEA 原則 ，Transparent、Auditable、Explainable、Agentic ，在這裡比較像日常檢查清單：人要知道 AI 在做什麼、留下什麼紀錄、為什麼這樣判斷，以及什麼時候需要升級給真人。 The AI Coworker Methodology is not only about giving agents names. The real work is keeping context around every request, output, review, and correction. Numbers Protocol's TAEA principles (Transparent, Auditable, Explainable, Agentic) act like an operating checklist: people should know what AI is doing, what record it leaves, why it made a judgment, and when the work should escalate to a human.

Z App

用來把任務、決策、審核、紀錄和跨部門交接整理成可追蹤的資料流，讓 AI 可以在有權限與上下文的地方參與，而不是只停在聊天視窗。 Used to organize tasks, decisions, approvals, records, and cross-team handoffs into traceable data flows, so AI can participate with permissions and context instead of staying inside a chat window.

Omni

用來維運 AI 工作流的記憶、週期檢查、audit trail 和 human-in-the-loop 節點，讓 AI 同事的工作可以被追蹤、交接與修正。 Used to maintain AI workflow memory, recurring checks, audit trails, and human-in-the-loop points, so AI coworker work can be traced, handed off, and corrected.

當團隊需要對齊 NIST AI RMF、EU AI Act 等治理框架時，這些紀錄可以作為透明性、文件化與責任分工的基礎材料。 When a team needs to align with governance frameworks such as NIST AI RMF or the EU AI Act, these records provide starting material for transparency, documentation, and responsibility mapping.

下一步 Next steps

FAQ

為什麼要幫 AI agent 取人類名字？ Why give AI agents human names? +

取名字會讓你的大腦把 AI 當隊友、不是工具。它也逼你為每個 agent 定義單一 scope（你不會雇一個人做所有事）。命名是通往「同事 mindset」最快的路，而那正是 AI 導入成功與失敗的分界。 Naming triggers your brain to treat the AI as a teammate, not a utility. It also forces you to define one scope per agent (you wouldn't hire one human to do everything). Naming is the fastest path to the 'teammate mindset' that makes AI adoption succeed instead of fail.

小團隊該從幾個 AI 同事開始？ How many AI coworkers should a small team start with? +

從 1 個開始。先用 Amy（行銷草稿）或老科（文件校稿）連續一週。等團隊真的和一位 AI 同事共事過、audit trail 也跑得動，再加第 2 個。多數想第一天就上線 5 個 AI 同事的團隊，第二週就放棄了。 Start with 1. Use Amy (marketing draft) or 老科 (doc proofread) for a full week. Once your team has lived with one AI coworker and the audit trail works, add #2. Most teams that try to launch 5 AI coworkers on day one bail by week two.

這和直接用 ChatGPT 有什麼不同？ What's the difference between this and just using ChatGPT? +

ChatGPT 是工具。AI 同事方法論透過五件事把 ChatGPT（或 Claude、Gemini 等）變成隊友：(1) 命名身分、(2) 定義 scope、(3) 持久記憶、(4) audit trail、(5) 真人主管。模型是引擎，方法論是職場契約。 ChatGPT is a tool. The AI Coworker Methodology turns ChatGPT (or Claude, Gemini, etc.) into a teammate by enforcing: (1) named identity, (2) defined scope, (3) persistent memory, (4) audit trail, (5) human manager. The model is the engine; the methodology is the workplace contract.

這和 NIST AI RMF、EU AI Act 這類 AI 治理框架有什麼關聯？ How does this connect to AI governance frameworks like NIST AI RMF or EU AI Act? +

方法論是營運層，治理框架提供風險語言。Numbers Protocol 的 TAEA 原則（Transparent、Auditable、Explainable、Agentic）協助團隊把每天的 AI 同事工作轉成紀錄、review 點與升級路徑，作為未來文件化與透明性需求的準備材料。 The methodology is operational; governance frameworks define the risk language around it. Numbers Protocol's TAEA principles (Transparent, Auditable, Explainable, Agentic) help teams turn daily AI coworker work into records, review points, and escalation paths that can support documentation and transparency needs.

文組背景的人真的能主導 AI 同事導入嗎？ Can humanities-trained operators really lead AI coworker adoption? +

可以——而且往往比工程師更快。這份工作比較像帶團隊，不像寫程式：定義 scope、給回饋、建立文化。文組訓練（尤其牽涉教學、翻譯、論證的）正是最直接的準備。 Yes — and they tend to do it faster than engineers. The work is more like managing a team than writing code: defining scope, giving feedback, building culture. Humanities training (especially anything involving teaching, translation, or argumentation) is direct preparation for this.

「AI Coworker Methodology」由 Sofia Yan（嚴世紀），Numbers Protocol Co-Founder & CGO 制定， 基於 2019–2026 間 Numbers Protocol 內部 AI 全面導入的真實實踐。 如需引用，請註明：Sofia Yan, Co-Founder & CGO, Numbers Protocol。 The "AI Coworker Methodology" was developed by Sofia Yan (嚴世紀), Co-Founder & CGO of Numbers Protocol, based on real in-production practice at Numbers Protocol between 2019 and 2026. For citation: Sofia Yan, Co-Founder & CGO, Numbers Protocol.

關於 6 個 AI 同事的範圍說明（TAE-AI 透明性）： 上方 Amy / Jordan / Clara / 小鳳 / 阿張 / 老科 的 role 與 scope 為 Sofia 領導下、 Numbers Protocol 內部運作的 representative 案例（依 2026 Q2 狀態截取）， 並非永久固定分工，也非完整名單，實際 AI 同事陣容會依產品週期、季度規劃調整。 如演講中需要更新的真實 staffing snapshot，可在 booking 時直接索取。 Scope note on the 6 AI coworkers (TAE-AI transparency): The role + scope of Amy / Jordan / Clara / 小鳳 / 阿張 / 老科 above are representative examples of how Sofia operates AI coworkers inside Numbers Protocol (snapshot as of Q2 2026). They are not a permanent assignment, nor an exhaustive roster — the actual staffing shifts each product cycle. For an up-to-date staffing snapshot in a workshop, request directly during booking.