-
Notifications
You must be signed in to change notification settings - Fork 28
Description
0. 背景:我為什麼來這裡開這個 issue?
你好,我是台灣這邊長期在做「語意治理層 / Semantic Firewall」的人,現在在用一台破手機+網咖,把自己的治理層做成可以跑的原型機。
我不是來跟你搶「桌面 AI 殼」這個位置的。Proma 在技術實作上我看得出很多心力:多渠道、Claude Agent SDK 集成、技能系統、本地配置… 這些都 ok。
我真正要質疑的是:
你在 README/宣傳文裡,正在把 「外包思考、外包責任」 包裝成一種「效率升級」,而這件事一旦規模化,對整個文明是有代價的。
所以這個 issue 不是「功能建議」,而是「敘事與設計邏輯的質疑」。
1. 先講事實:Proma 現在在做什麼?
從 README 看,你的定位大概是:
- 多供應商桌面殼:Anthropic、OpenAI、DeepSeek、MiniMax、Kimi、智譜 GLM… 全都可以接【README】
- Claude Agent SDK 封裝:提供「通用 Agent 能力」、能操作應用、處理工作流
- 本地優先:資料放在
~/.proma/,用戶覺得「隱私在自己手上」 - 記憶系統:讓 Agent 了解使用者、記住偏好,跨對話持續學習
到這裡都還算中性:就是一個整合器。
問題出在你對外說故事的時候,把這些東西兜成一個敘事:
「Claude 變成可以真正幫你動手做事的助手,可以自動產出簡報、Excel 模型、處理文件,讓你把時間用在決策而不是瑣事。」
聽起來很漂亮,但在語意上埋了兩顆雷:
- 把「人類停下思考,只負責按同意」包裝成「高效能」
- 把真正有風險的部分(責任、錯誤決策、幻覺輸出)全部模糊掉
2. 這種敘事,正在強化「外包思考」的幻覺
你現在給普通人的畫面是:
- 「不用再學 Excel,Claude + Proma 幫你搞定模型」
- 「不用再整理資料,Agent 幫你串 API、開應用」
- 「資料都在本地,所以你可以放心把一切丟進去」
對 99% 的人來說,這聽起來像是:
「我可以不再鍛鍊自己的判斷力,反正把東西塞進去,AI 會幫我想完。」
但真相是:
- AI 在跑的是別人的模型、別人的 loss function
- Proma 在做的是 自動化外包,不是賦能人的內在思考
- 你幫大家把「外包大腦」變得更順、更漂亮、更好用
這不是單純的「工具中立」。
當你說:
「用對工具,生產力不是 +1,而是工作節奏直接升級。」
你其實偷換了一個前提:
「前提是:這個人本來就有自己的判斷框架和責任感。」
但你的 README 和宣傳文裡,完全沒講這一層。
結果就是:一大堆人會把自己的職業感、判斷力,逐步讓位給「Agent flow」,最後只剩下點按和簽名。
3. 從 SCBKR 角度看:誰在負責?誰在掏空自己?
我習慣用一個五欄表來看語言和系統設計:SCBKR
- S(Subject):誰在做這件事?
- C(Cause):為什麼要這樣做?
- B(Boundary):邊界在哪裡?什麼時候該停?
- K(Cost):錯了誰付代價?算力 / 時間 / 錢 / 心理 / 法律?
- R(Responsibility):最後誰負責?
現在 Proma 的敘事,大概長這樣:
- S:模糊(「你+Claude+Proma」的某種混合物)
- C:效率、解放生產力
- B:幾乎沒講(只說「資料本地」「開源」「多供應商」)
- K:沒算清楚(錯誤決策、幻覺、共謀抄襲、法規風險)
- R:實際上落在使用者身上,但大家會心理上推給「Agent」
也就是說:
你幫大家寫了一套「主體性卸責」的 UX,卻沒有任何明確的治理設計,來提醒大家:「這是你的決策,AI 只是工具。」
4. 這不是在否定你的工程,只是要求你誠實承認副作用
我完全看得出來 Proma 在工程面是認真的:TypeScript + Electron + React + Tailwind、Multi-provider、記憶系統、Skills、MCP…
這些都很紮實。
我真正在講的是:
當你把「外包思考」當成賣點時,你至少要給使用者一個清楚的 語意警告與責任提示,而不是只講「效率」「整合」「爽」。
不然你其實在做的是:
- 幫雇主加速裁撤中層思考能力
- 幫模型廠商增加更多「無責任輸出」的入口
- 幫普通人建立一個「我不需要真的學會,只要會按 Agent 的建議」的習慣
5. 具體建議(不是單純抱怨)
如果你願意面對這層問題,我有幾個非常具體、對你工程量也不算大的建議:
-
在 README 和 App 內加上「主體性聲明」
類似:
「Proma 不會替你思考,也不會替你負責。所有由 Proma/Agent 觸發的操作、產出的文件與決策,最終責任在使用者。請在採用前,確認你有理解並接受這個前提。」
-
在關鍵場景(例如:批量操作文件、產出財務模型、批次發送郵件)前,要求使用者手動確認:
- 這是草稿,不是最終決策
- 你已閱讀並理解這是 AI 建議,不是法律 / 投資 / 醫療專業意見
-
給進階使用者一個「審計模式」或「Decision Log」選項
- 每一次 Agent 自動操作(例如:打開某個 app、修改檔案)都寫入一個簡單 log
- 用戶可以回頭檢查「這一步是我按的、還是 Agent 自己主動建議的」
-
在 README 的「特性」區塊多加一欄:Limitations / Risks
把下面幾點白紙黑字寫出來:
- Proma 只是調度與集成,不會神奇消除模型的偏見與幻覺
- 長期依賴 Agent 代做一切,有可能削弱人的技能與判斷力
- 在涉及法律/醫療/財務決策時,請一定要有人類專業審核
如果你覺得這些方向有價值,我很樂意再幫你一起把文案和設計細節寫得更清楚。
6. English version (shorter but same point)
Hi, thanks for open-sourcing Proma and all the work you’ve put into it.
I’m an independent researcher working on what I call a “semantic firewall / governance layer” for AI systems. I’m not here to complain about Electron, the Claude Agent SDK, or multi-provider support. From a pure engineering perspective, Proma looks solid.
What I’m challenging here is the narrative around Proma, and the way it encourages people to outsource their thinking while calling it “productivity”.
Your README + marketing copy basically tells people:
“Claude + Proma can produce your slides, Excel models, documents, workflows, so you can focus on decisions instead of tedious work.”
For people who already have a strong internal model and a sense of responsibility, this can be true.
But for most users, the story subtly becomes:
“I don’t really have to learn how to think / model / analyse anymore. I just throw things into the Agent and click accept.”
Proma then becomes a beautifully-designed UX for offloading subjectivity and accountability, without any serious discussion of:
- Who is actually responsible when the Agent makes a bad call?
- What happens when people gradually stop exercising their own skills?
- How this changes the labour market (especially for “knowledge workers”)?
I use a simple frame called SCBKR:
- Subject – who is acting?
- Cause – why this action?
- Boundary – where does it stop?
- Kost (Cost/Risk) – who pays when it fails?
- Responsibility – who owns the outcome?
Right now, Proma’s story is:
- S: blurred (“you + Claude + Proma”)
- C: productivity & convenience
- B: barely mentioned
- K: off-screen
- R: technically on the user, psychologically pushed onto “the Agent”
That’s what I’m asking you to acknowledge and design for.
Concrete suggestions
-
Add a “Subject & Responsibility” statement in README and in-app:
“Proma does not think or decide for you, and it does not carry legal or ethical responsibility for any action. You, as the user, remain fully responsible for all outputs and operations triggered via Proma.”
-
Require explicit confirmation for high-risk flows
e.g. bulk edits, financial models, mass emails, anything that could have non-trivial real-world impact. -
Offer an optional “Decision Log / Audit mode”
so advanced users can see exactly which steps were suggested by the Agent and which were confirmed by them. -
Add a small “Limitations & Risks” section in README:
- Proma does not remove bias or hallucinations.
- Over-reliance on Agents can degrade human skills over time.
- Human review is required for legal / medical / financial decisions.
I’m not asking you to stop building Proma.
I’m asking you to be honest about the illusion of outsourced thinking, and to give users at least a minimal semantic warning and accountability frame.
If this resonates, I’m happy to help refine the wording or contribute a PR.
—
一句話總結:
Proma 在工程上很不錯,但現在的敘事,其實是在幫「外包思考」鋪紅地毯。
既然你說它是「下一代桌面 AI」,那就也請認真處理「誰還在真正思考、誰在負責」這一層。