Ciao i'm ciaO Product designer (AI) · Taipei
← back to work← 回到作品
Case Study 02 · TU/e Eindhoven · 2025案例 02 · TU/e 安荷芬 · 2025

Creative Gizmo, a collaborative AI agent that blends into the brainstormCreative Gizmo,融入腦力激盪的協作型 AI agent

A working prototype, a Level of Involvement framework, and six guidelines for AI that collaborates without taking over.一個可運作的原型、一套 Level of Involvement 框架,和六條「會協作、不喧賓奪主」的 AI 設計準則。

Creative Gizmo cover: a shared whiteboard tablet with three color-coded AI modes, while the team says let's ask Gizmo for help
Role角色
Prompt layer + UX researchexperiment design · interviews · analysisprompt 底層 + UX 研究實驗設計 · 訪談 · 問卷分析
Duration時間
~5 monthsSep 2025 to Jan 2026約 5 個月2025 年 9 月到 2026 年 1 月
Tools工具
TLDraw · Whisper · ChatGPT (GPT-4.1/5.1)UI vibe-coded with GPTTLDraw · Whisper · ChatGPT (GPT-4.1/5.1)介面用 GPT vibe coding
Team團隊
Team of 4 · TU/etechnical implementation · UX design · product design四人團隊 · TU/e技術實現 · UX 設計 · 產品設計

01 Background背景

Most creative AI tools are built to automate. Their heavy, dominant output drowns out the user's own thinking. We wanted an AI that joins a brainstorm as a collaborator, without taking ownership of the ideas.多數創意類 AI 以自動化為導向,強勢又大量的輸出會蓋過使用者自己的思考。我們想讓 AI 以協作者的身分加入腦力激盪,而不奪走點子的主導權。

02 Problem問題

From market research, three problems stood out:市場調查後,我們發現三個問題:

Automation-first AI floods the board and overshadows the people
📣AI overshadows the user's thinkingAI 蓋過了使用者的思考

It floods the board before people can organize their own thoughts, turning creators into passive receivers.使用者還來不及整理想法,白板就被 AI 灌滿,從創作者變成被動接收者。

AI interrupts at the wrong moment and breaks the group's flow
⏱️AI can't read the rhythm of the conversationAI 抓不到對話的節奏

It can't follow the live discussion, so it cuts in at the wrong moment.跟不上現場的討論,常在錯的時機插話。

A tug-of-war over who holds authority of the ideas
🌱AI should extend ideas, not take the wheelAI 該延伸想法,而不是搶主導權

Most tools ignore where the discussion is and keep generating on their own.多數工具無視討論的進展,自顧自地不斷生成。

The most useful AI isn't the one that says the most, it's the one that knows when not to.最有用的 AI,不是說得最多的,而是知道何時不說的。

03 Challenge挑戰

How might we let AI take part in a live, co-located brainstorm, helpful enough to move thinking forward, yet bounded enough that the team always keeps ownership of the ideas?我們如何讓 AI 參與一場面對面的即時腦力激盪,幫得上忙、能推進思考,又有分寸到讓團隊始終握有點子的主導權?

04 Research & Insights研究與洞察

Our research ran in two layers: exploration, then validation.研究分兩層:先探索,再驗證。

Exploratory experiment探索實驗
Brainstormed with a voice AI與語音 AI 腦力激盪
Noting where the AI helped or got in the way, organized into do's and don'ts記下 AI 幫上忙與造成干擾的地方,整理成 do's and don'ts
Sketches and notes from brainstorming with a voice AI
Concept explorations: physical devices and mechanisms to limit the AI
Concept development概念發展
Explored multiple concepts多個概念探索
Physical devices, mechanisms to limit the AI實體裝置、限制 AI 的機制等
Wizard-of-Oz studyWizard-of-Oz 研究
A human simulating the AI facilitator真人模擬 AI facilitator
One team of four, without vs with AI, observation + interviews一組四人,無 AI vs 有 AI 對照,觀察 + 訪談
The Wizard-of-Oz session: participants brainstorming while a camera records
Do's and don'ts of AI participation from the exploratory experiment
Design guidelines設計準則
Distilled into guidelines收斂成設計準則
Findings + limits → guidelines發現 + 侷限 → 準則
Working prototype可運作原型
Built on TLDraw建在 TLDraw 上
Mode prompts iterated v01 → v02模式 prompt 迭代 v01 → v02
The working prototype built on TLDraw
User testing session with the working prototype
User testing使用者測試
Tested in real brainstorms真實腦力激盪實測
Three participants, observation + a nine-point questionnaire三人實測,觀察 + 九點量表問卷
Don'ts · issuesDon'ts · 出問題的
What the AI didAI 做了什麼
Overly dominant過度強勢Too much talking話太多AI-centred以 AI 為中心Unprompted intervention未經邀請就介入Too complex太複雜
What it caused團隊因此怎樣
User dependency使用者依賴Unequal participation參與不平等Reduced collaboration協作變少
Do's · worked wellDo's · 有效的
When to step in什麼時候出手
On-demand support被請求才支援Technical clarification技術澄清Knowledge source知識來源
How to help用什麼方式幫
Asking questions主動發問Making connections建立連結Providing structure提供結構Stimulating reflection激發反思
Do's and don'ts from the exploratory experiment.探索實驗整理出的 do's and don'ts。

Legend: colored blocks are discussion phases; the blue marker is when the AI was invited to ask a question.圖例:色塊是討論階段;藍色標記是 AI 被邀請提問的時刻。

(A)The brainstorm without AI無 AI 的腦力激盪過程

Topic: waste disposal in a residential building題目:住宅大樓的垃圾處理
Personal experience個人經驗Hometown dryer
Singapore garbage chute
A fridge for food waste
Insect composting bin
家鄉的烘乾機
新加坡垃圾滑道
存廚餘的冰箱
昆蟲堆肥箱
Making decisions做決定Summarize possible forms
Limited conditions
整理可能的形式
限制條件
Problems問題Waste stays
Going downstairs
垃圾滯留
要下樓丟
Solutions解法An organization in the building
A communal space
Slide or rope
Redesign the kitchen system
大樓內的組織
公共空間
滑道/繩索
改造廚房系統
Limits侷限
Decisions做決定
13579111315 min
Hover or tap a phase to see the ideas raised.滑鼠移到色塊上或點一下,查看當時提出的想法。

(B)The brainstorm with AI有 AI 的腦力激盪過程

Topic: supporting study focus with music題目:用音樂輔助學習專注
Question from AIAI 提問
Understand context & problems理解脈絡與問題Getting into flow, how to start
Depends on the task, how to stop
Attention decreasing
進入心流、如何開始
取決於任務、如何停下
注意力下降
Personal examples個人例子Different playlists suit different learning activities不同歌單適合不同學習活動
An idea一個點子Music reflects the state of your study音樂反映學習狀態
Decide to ask AI決定問 AIChoose AI選擇 AI
Converging around the AI's question圍繞 AI 的提問收斂出解法Break reminder → design detail
Connect with health parameters
Sense movement, incorporate cameras
休息提醒 → 設計細節
連結健康參數
感測動作、結合鏡頭
13579111315 min
Hover or tap a phase to see the ideas raised.滑鼠移到色塊上或點一下,查看當時提出的想法。

Without AI, ideas diverged and struggled to converge. With AI, the team converged faster, and only called on it when stuck, preferring questions over direct answers.無 AI 時想法發散、難收斂。有 AI 時收斂變快,而且大家只在卡住時才求助,偏好提問勝過直接給答案。

Six core guidelines for collaborative AI協作型 AI 的六條核心準則

Listen continuously, act cautiously持續聆聽,謹慎行動
Finding發現Talk drives the brainstorm, sticky notes come second討論主要靠口說,便利貼其次
Limit侷限Tone and body language go uncaptured語氣、肢體語言等線索抓不到
Intervention starts with the user介入由使用者發起
Finding發現People want AI only at the right moment大家偏好 AI 在對的時機才介入
Limit侷限Team dynamics stay invisible to the system團隊裡的互動關係,系統看不見
Ask questions by default預設以提問引導
Finding發現Questions support without taking over提問能支援討論,不搶主導
Limit侷限Understanding depends heavily on contextAI 的理解高度依賴脈絡
Keep direct ideas as an option保留直接給點子的選項
Finding發現Direct ideas are wanted only when stuck只有卡住時,才想要 AI 給點子
Limit侷限The setup may overstate the AI's presence模擬實驗可能高估 AI 的存在感
Few options, clear modes選項要少,模式要清楚
Finding發現Teams shift between diverging and converging團隊自然在發散與收斂間切換
Limit侷限More modes, more learning cost模式一多,就有學習成本
Regulate by timing and form, not quantity用時機與形式節制,而非數量
Finding發現People used the AI sparingly throughout大家全程節制地用 AI
Limit侷限Short sessions say little about long-term use場次太短,看不出長期樣貌

05 Solution解法

Creative Gizmo is a shared whiteboard built on TLDraw, with an AI that acts like an invited teammate. Through the Level of Involvement framework, the team, not the AI, decides how much it joins in, by switching between three modes.Creative Gizmo 是建在 TLDraw 上的共享白板,裡面的 AI 像一位被邀請的隊友。透過 Level of Involvement 框架,由團隊而不是 AI 決定它參與多少,方法是在三種模式間切換。

Automated
What underlying qualities or experiences do you associate with a 'premium' product, and how might those guide your decisions about the design and materials for the automated cat feeder?
RESERVED
Only asks questions, never touches the board只提問,不動白板
Automated
Here's a thought: imagine the automated cat feeder as a luxury watch for pets, just as a fine watch combines precision engineering, premium materials, and emotional value, your feeder could blend technology, elegant design, and a sense of care.
ENGAGING
Widens the discussion with analogies用類比拓寬討論
Automated
Cat feeder
Premium materials: glass, ceramic, metal?
What questions do we want to solve?
How can we make it eco-friendly?
Three post-its encourage the group to explore premium materials, focus on emotional impact, and consider sustainability.
ACTIVE
Adds idea post-its directly直接貼出便利貼

I designed and tuned the prompt layer: one dedicated prompt per mode, each with a clear role and exact output caps.底層的 prompt 由我設計與調校:每個模式一個獨立 prompt,各有明確角色和精確的輸出上限。

Whisper turns the live audio into a running meeting summary. When someone presses the facilitator button, the current mode's prompt, the summary and a whiteboard screenshot go out as one request. That keeps the AI's behavior stable and predictable, which users valued most.Whisper 把現場對話轉成持續更新的會議摘要。有人按下 facilitator 按鈕時,系統把當前模式的 prompt、摘要和白板截圖打包成一次請求。AI 的行為因此穩定、可預期,這正是使用者最看重的。

Full system design: audio recording through WhisperAI, conversation summarization, the three mode prompts behind the facilitator button, and the whiteboard output
The full system design.完整系統設計。

06 Design decisions設計決策

human avatar → abstract orb
🫧From a human avatar to an abstract orb化身:從擬人頭像到抽象球體

Testers said the human avatar pulled too much attention, so we replaced it with an abstract orb that breathes.受測者反映擬人頭像太搶注意力,所以改成一顆會呼吸的抽象球體。

after user testing ↓使用者測試後 ↓
RESERVED
ENGAGING
ACTIVE
🎨Color-coding the modes, driven by testing測試回饋:模式的顏色編碼

The three modes were hard to tell apart in testing, so we color-coded them by arousal level: calm blue, lively yellow, high-arousal red. The palette is not user-validated yet.測試中三種模式很難分辨,所以我們依喚起度上色:冷靜的藍、活潑的黃、高喚起的紅。這套配色尚未經過使用者驗證。

07 Outcomes成果

Demo day: the team presenting Creative Gizmo, walking visitors through the system, and sticky-note feedback left on the Gizmo poster
Demo day: presenting Creative Gizmo to visitors.期末展示:向觀展者介紹 Creative Gizmo。

In testing with three participants, question-based support and automatic summaries were valued most. Predictability and transparency made people comfortable keeping the AI at the table.三位參與者的實測中,提問式支援和自動摘要最被看重。可預期、透明,是大家願意把 AI 留在桌上的關鍵。

We delivered three things: a working prototype, a reusable Level of Involvement framework, and six design guidelines.最後交出三樣東西:可運作的原型、可複用的 Level of Involvement 框架,和六條設計準則。

08 Research limitations研究限制

Small, mixed sample樣本小、組成不一

Four people in the Wizard-of-Oz study, three in the prototype test.WoZ 四人、原型實測三人。

nextA larger, more consistent sample, followed over time.更大、更同質的樣本,並長期追蹤。
Latency breaks the flowAI 延遲打斷心流

While waiting for the AI, the discussion often paused.等 AI 回應時,討論常會停下來。

nextA faster loop, or a wait that keeps people engaged.更快的互動迴圈,或讓等待保持參與感。
Mode switching drops context切換模式會丟失脈絡

Chat history is cleared between modes to prevent role bleed.切換模式會清空對話歷史,以避免角色互相汙染。

nextA main agent that keeps the context and orchestrates the three modes.由一個主 agent 保留脈絡、統籌三種模式。
← all work← 所有作品 next: Smart Coastline →下一個:Smart Coastline → say ciao →說聲 ciao →