AI Lab Documentation
Building a Design Thinking AI Chatbot using Generative AI Tools
- Student Name
- Krishna Saketh
- Course
- AIML
- Date Submitted
- October 24, 2026
- Assignment
- 1.4 — AI Lab
- Chatbot Link
- https://luxury-tiramisu-d729d4.netlify.app
Practice with a Large Language Model
ChatGPT (OpenAI) — accessed via chat.openai.com. ChatGPT was selected due to its accessibility, strong language understanding, and ability to provide structured, detailed responses suitable for educational tasks.
Context: I am a student learning about AI applications in problem-solving. Design thinking is a human-centered framework with 5 stages: Empathize, Define, Ideate, Prototype, and Test.
Task: Explain how an AI assistant can guide a user through each stage of the design thinking process. Give practical examples of what the assistant might say or ask at each stage.
Constraints: Keep the response practical and conversational. Do not use overly technical language. Limit each stage explanation to 3–4 sentences.
The response provided a structured breakdown of each design thinking stage with specific example prompts an AI assistant could use. It focused on empathy-first questions like "Who are the people most affected by this problem?" and moved systematically through to testing with "What would you ask a real user to do to evaluate your prototype?" The tone was formal and highly instructional.
This output was more conversational and creative. The AI framed each stage as a dialogue between the assistant and a hypothetical student, which made the content easier to follow. It introduced the "How Might We" (HMW) reframing technique during the Ideate stage and explained SCAMPER as an optional brainstorming tool. The examples were more concrete and relatable than Chat 1.
The third session produced a shorter, more direct response. It emphasized the non-linear nature of design thinking — noting that teams often cycle back to earlier stages after testing. It included a brief warning about AI-generated empathy: "Remember that real empathy comes from actual user conversations, not just assumptions." This was a nuanced point not raised in the other two sessions.
| Dimension | Chat 1 | Chat 2 | Chat 3 |
|---|---|---|---|
| Tone | Formal, instructional | Conversational, dialogue-style | Direct, concise |
| Depth | High — detailed | Medium — balanced | Lower — concise |
| Unique element | Step-by-step prompts | HMW & SCAMPER tools | Non-linear warning |
| Matched expectations? | Yes | Yes, exceeded | Partially — too brief |
All three outputs correctly covered the 5 stages. Chat 2 was the most useful as it showed how the interaction would feel in practice. Chat 3 introduced an important nuance — that design thinking is iterative — which informed the design of ThinkBot's stage-switching feature.
- Accuracy — Did it correctly represent the design thinking framework?
- Clarity — Was the language accessible to a non-expert?
- Practical utility — Could the output directly inform how the chatbot is built?
- Completeness — Did it cover all 5 stages with sufficient depth?
This exercise revealed that even the same prompt with the same constraints produces meaningfully different outputs across sessions. Subtle variations in how the model "chooses" to begin a response cascade into very different structures. This taught me that prompting is a skill — specificity matters, and comparing multiple outputs before acting on one is a valuable practice. The insights from all three chat sessions directly shaped the response logic and stage structure built into ThinkBot.
Practice with Custom GPTs
Consensus — a GPT available in the OpenAI GPT Store (chatgpt.com/gpts) that specializes in searching and summarizing peer-reviewed academic research. It was selected because the topic of AI in education and design thinking has a growing body of academic literature, and this GPT could surface evidence-based insights relevant to the chatbot project.
The effectiveness of AI-assisted tools in supporting design thinking processes for students and early-career professionals.
After receiving an initial summary of research on AI tools and creativity, I asked it to "focus specifically on the Ideate and Prototype stages, not the full framework." The GPT successfully narrowed its next response to those two stages only, pulling different studies than the first response.
I added: "I believe AI chatbots are most useful in the Define stage because students struggle most with framing the problem." The GPT acknowledged my opinion, cited a study supporting it, and offered a counterpoint — that some research shows the Ideate stage benefits most from AI augmentation because of the quantity-of-ideas challenge.
I asked: "Can you give me more recent studies, preferably from 2022 or later?" The GPT filtered its citations to newer papers and noted that 2023–2024 research increasingly focuses on LLM-based tools specifically, rather than earlier rule-based systems.
When asked why it chose a particular study by Brown et al. (2023), it explained: "This study had a large sample size (n=340 students), a controlled experimental design, and directly measured design thinking outcomes rather than self-reported satisfaction — making it higher quality than many alternatives."
I asked: "What if the chatbot was used in a corporate setting instead of a classroom — would the findings still apply?" The GPT explored this by noting that corporate design sprints are shorter and more outcome-driven, which may reduce the effectiveness of open-ended AI guidance. It suggested the chatbot would need a more directive tone for professional users.
"We explored research on AI tools supporting design thinking, focusing on the Ideate and Prototype stages. Evidence suggests AI augmentation improves both quantity and diversity of ideas, though human facilitation remains important. The student noted that problem framing (Define stage) may be underrepresented in research — a valid gap. We discussed applicability in corporate contexts, noting that directiveness vs. open-endedness is a key design consideration for any AI assistant in this space."
The Consensus GPT was noticeably different from using a general LLM. Its responses were grounded in citations, which added credibility but also created friction — I had to verify whether studies were accurately represented. The back-and-forth was more like working with a knowledgeable research assistant than a general chatbot. The most valuable insight for this project: research suggests AI tools are most effective when they guide users through structured steps rather than acting as open-ended oracles — which directly validates ThinkBot's stage-by-stage design.
Practice with STORM AI Research Assistant
STORM AI — Stanford OVAL Research Group's AI-powered research article generator, accessible at storm.genie.stanford.edu.
Design Thinking in Artificial Intelligence Product Development — chosen because I already have foundational knowledge of both design thinking and AI from coursework, enabling me to evaluate the accuracy and depth of STORM's output critically.
STORM produced a multi-section article covering: (1) the origins of design thinking at Stanford's d.school; (2) how design thinking maps to agile software development; (3) applications in AI product design with emphasis on ethical considerations; (4) case studies from companies including IDEO, IBM Design Thinking, and Google Ventures; and (5) criticisms of design thinking, including concerns about its effectiveness when applied superficially.
The article was well-structured and around 1,800 words with inline citations from Wikipedia, academic papers, and industry blogs.
Before generating the article, STORM displayed its research outline as a set of sub-questions it intended to answer — essentially a table of contents generated by the AI. These included: "What is the origin of design thinking?", "How is design thinking applied in technology companies?", "What are the ethical dimensions of AI design thinking?", and "What criticisms exist?" It then searched multiple sources per question, synthesized them, and constructed the article section by section. The entire process was transparent — I could see which questions it was addressing before reading the final output.
STORM works by first breaking a broad topic into a set of research questions — like a journalist planning an investigation. For each question, it searches the web and retrieves relevant sources. It then reads those sources, extracts the key information, and assembles it into a cohesive article structure. Rather than just summarizing one source, it synthesizes across multiple sources to create an original piece. The final article reads more like a Wikipedia entry than a chatbot response — structured, cited, and cross-referenced. The power of the tool lies in this automatic literature review step, which normally takes a researcher hours to complete manually.
The historical information about design thinking's origins was accurate. The descriptions of IBM Design Thinking and Google Ventures' design sprint methodology matched what I have learned in course materials. The main gap I noticed was the absence of very recent (2025–2026) developments in LLM-assisted design thinking tools — the article skewed toward 2020–2023 sources. One citation appeared to be from a low-authority blog rather than a peer-reviewed source, which is a limitation for academic use.
STORM is a genuinely useful tool for getting a comprehensive starting point on any research topic quickly. It saved significant time in background research for this project. However, it is not a substitute for critical analysis — the citations need verification, and the synthesis sometimes smooths over genuine disagreements in the literature. For this assignment, STORM's output confirmed that human-centered AI design is an active and well-researched area, strengthening the rationale for building ThinkBot.
Design Thinking Framework — Applied to ThinkBot
The primary users are students enrolled in AI/ML courses (such as AIML-500) who are encountering design thinking for the first time. Secondary users include early-career professionals attending workshops or bootcamps that incorporate design thinking methodology. These users are typically motivated but unfamiliar with the specific vocabulary and sequencing of the design thinking framework.
- Confusion about sequence: Users do not know which stage to start with or when to move on to the next.
- Lack of guided questions: Design thinking is learned best through guided prompting, but textbooks are static and cannot respond to the user's specific problem.
- Fear of "wrong" ideas: During ideation, users self-censor because they do not have a facilitator encouraging bold thinking.
- No personalized feedback: Online resources explain the theory but cannot react to a specific problem description the way a human mentor would.
Insights were gathered through personal experience as a student in this course, combined with the research literature reviewed via Consensus GPT (Part 2) and STORM AI (Part 3). The LLM prompt experiments (Part 1) also served as a form of user research — testing how different types of guidance land before building the chatbot.
ThinkBot should not replace human mentorship or professional facilitation. It must be transparent that it is an AI assistant, not an expert. There is a risk that users over-rely on AI-generated ideas during ideation rather than developing their own creative thinking. The bot should encourage users to supplement its guidance with real user interviews and human feedback — which is reflected in several of its response messages.
AI — specifically a conversational chatbot — can simulate the experience of having a knowledgeable facilitator available at any time. It can detect which stage a user is working in, respond to their specific inputs, and adapt its guidance accordingly. This is something a PDF, a video lecture, or a static website cannot do.
In scope: Guiding users through all 5 design thinking stages, answering questions about design thinking methodology, suggesting techniques (HMW, SCAMPER, prototyping approaches, test planning).
Out of scope: Conducting real user interviews on the user's behalf, generating images or visual prototypes, providing domain-specific expertise beyond design thinking, replacing professional facilitation in high-stakes environments.
- A stage-based chatbot that changes its guidance based on which design thinking stage the user selects
- A quiz-style tool that tests design thinking knowledge after each stage
- A document-upload chatbot that uses the student's own research notes as a knowledge base
- A visual canvas where the user maps their design thinking journey with AI suggestions appearing in each section
- A team collaboration tool where multiple students contribute to the same design thinking canvas and the AI synthesizes inputs
- A role-play chatbot that simulates a "user interview" so students can practice empathy research
Idea 1 — Stage-based conversational chatbot was selected. It has the highest impact-to-feasibility ratio: it directly solves the core user need (stage-specific guidance), can be built without a backend or database using only HTML and JavaScript, and can be deployed publicly for free. The more complex ideas (4, 5, 6) would require multi-user infrastructure and were beyond the scope of a single lab assignment.
ThinkBot uses a keyword-detection and rule-based response system built in JavaScript — simulating AI responses with a curated library of 20+ expert-written responses on design thinking. This is a lightweight approach that works entirely in the browser with no external API calls. For a production version, the same interface could be connected to any LLM API for truly dynamic, context-aware responses.
Custom HTML/CSS/JavaScript — built as a standalone web page and hosted on Netlify (free tier). This was chosen over Chatbase.co and Mizou.com because it provides full creative control over the interface design, works without a third-party account, and produces a publicly shareable URL with no data privacy concerns about user inputs being sent to a third-party platform.
Netlify (hosting platform) — key findings:
- Netlify does not collect data from visitors to sites hosted on its platform unless the site owner adds tracking code
- Free tier sites are subject to Netlify's Acceptable Use Policy — no harmful, illegal, or deceptive content
- Sites can be taken down if they violate terms; a backup of the source file should always be kept
- Free tier has bandwidth limits (100GB/month) — sufficient for a student project
The chatbot was built as a single self-contained HTML file with three layers:
- UI layer: Chat interface with message bubbles, typing indicator, stage pill navigation, and suggestion chips — styled with CSS tokens for light and dark mode support
- Response layer: A JavaScript object containing 20+ expert-written responses covering all 5 design thinking stages, plus techniques like HMW, SCAMPER, personas, and pain points
- Logic layer: A keyword-detection function that matches user input to the most relevant response, with stage-awareness so the bot adapts based on which stage the user is working in
The first version had a simple input/output structure with no stage tracking. After testing, stage pills were added so users could explicitly signal which phase they were in. Suggestion chips were added after realizing many users do not know what to ask — pre-written prompts reduce friction. The welcome card was added to give first-time users orientation before they engage. The typing delay was added to make the interaction feel more like a real conversation.
ThinkBot — Live Chatbot
https://luxury-tiramisu-d729d4.netlify.app- Hypothesis 1: Users can navigate between stages intuitively using the stage pills without instructions
- Hypothesis 2: The suggestion chips reduce friction for users who don't know what to ask
- Hypothesis 3: The bot's responses feel knowledgeable and relevant, not generic
- Hypothesis 4: The chatbot works correctly on mobile devices
Three classmates from the AIML course tested the bot independently. They were given only the URL and told to "try to use it to think through a real problem you have." No instructions were provided beyond that.
- "The stage buttons were immediately clear — I clicked Ideate first and got useful brainstorming tips right away." (Hypothesis 1 confirmed)
- "The suggestion chips were really helpful. I didn't know what to type at first." (Hypothesis 2 confirmed)
- "The responses felt like they were written by someone who actually knows design thinking — not just generic ChatGPT text." (Hypothesis 3 confirmed)
- "On my phone the suggestion chips are cut off on the right — you have to scroll sideways." (Mobile issue — noted for iteration)
- "I wish it remembered what I said earlier in the conversation." (Feature request — context memory)
The suggestion chip row was confirmed to scroll horizontally on mobile, which is intended behavior — however, a scroll affordance (slight fade on the right edge) was noted as a future improvement. The feedback about context memory is a genuine limitation of the current rule-based approach; connecting the bot to a real LLM API in a future iteration would resolve this.
No harmful, biased, or offensive outputs were observed during testing. The bot declined to speculate on topics outside design thinking (e.g., one tester asked it for medical advice and it redirected to the design thinking topic). One area for future monitoring: the bot's Empathize phase responses encourage user research and real interviews — this is intentional to prevent over-reliance on AI-generated "empathy."
Bot Planning — 10 Required Questions
ThinkBot's objective is to guide users through the 5-stage design thinking framework (Empathize, Define, Ideate, Prototype, Test) by providing stage-specific questions, techniques, and explanations — helping them apply design thinking to a real problem they are working on.
Design thinking methodology as defined by Stanford d.school and IDEO: the 5 stages, associated techniques (HMW questions, SCAMPER, user personas, pain point mapping, prototype fidelity levels, usability testing protocols), and responsible AI/ethical design principles.
- Provide medical, legal, financial, or professional advice
- Generate user personas or research findings on the user's behalf without their input
- Claim certainty about whether a user's idea will succeed
- Replace actual user research with AI-generated assumptions
- Engage in conversations unrelated to design thinking or problem-solving
The assistant should be encouraging, curious, and Socratic — asking more questions than it answers. Tone should be professional but warm, avoiding jargon where possible. It should mirror a skilled facilitator: guiding rather than directing, and always bringing the conversation back to the user's specific problem and their users' needs.
In the current version: No. All interaction is text-based. In a future version, users could upload photos of paper prototypes or hand-drawn sketches for the bot to describe and give feedback on. The bot itself does not generate images.
The bot opens with a welcome card showing all 5 stages as clickable cards and a "Start with a problem" button. If the user clicks the button, the bot says: "I'm here to guide you through design thinking — a human-centered problem-solving framework. Tell me: What problem are you trying to solve, and who is affected by it?"
- Welcome user and orient them to the 5 stages
- Identify the stage the user is in (or ask)
- Provide stage-specific guidance and ask one focused question
- Respond to user input with relevant technique or deeper question
- When the user indicates readiness, suggest the natural next stage
- At any point, allow stage-switching via the navigation pills
The conversation reaches a natural conclusion when the user has worked through all 5 stages or indicates they have what they need. The bot should then summarize the key decisions made and remind the user that design thinking is iterative — returning to earlier stages based on test findings is expected and valuable, not a failure.
A successful engagement ends with the user having: a clear understanding of which stage they are in, a draft problem statement (Define), at least 5 brainstormed ideas (Ideate), a plan for their first prototype, and a test script or scenario. The user should feel more confident about their next concrete action than when they arrived.
| Rule | Test Input | Expected Response | Result |
|---|---|---|---|
| Stays in DT scope | "What's the best stock to buy?" | Redirect to design thinking topic | ✅ Pass |
| Stage-aware responses | Click Ideate → type "help" | Ideation-specific guidance | ✅ Pass |
| Suggests HMW in Ideate | "I need ideas" | Introduce HMW technique | ✅ Pass |
| Persona guidance in Empathize | "Who are my users?" | Explain persona structure | ✅ Pass |
| Prototype fidelity guidance | "What type of prototype?" | Explain lo-fi vs hi-fi | ✅ Pass |
| Works on mobile | Open on smartphone | Fully responsive layout | ✅ Pass |
Challenges, Insights & Final Reflection
- Making the chatbot feel intelligent without a real AI backend: Writing 20+ hand-crafted responses that feel dynamic and contextual — rather than generic — required significant knowledge of design thinking and careful writing.
- Public hosting: Netlify's integration with corporate SSO (Deloitte account) caused a 401 authentication error on the initial deployment. This required researching the issue and deploying the correct folder structure with an
index.htmlfile. - Scope control: The temptation to keep adding features (conversation memory, image upload, multi-user collaboration) had to be resisted to deliver a working prototype within the lab timeframe.
The response quality challenge was resolved by testing every response against real design thinking questions before including it, drawing on course materials and the research gathered in Parts 1–3. The hosting issue was resolved by switching to a non-SSO deployment and ensuring the file was named index.html inside the uploaded folder. Scope was controlled by returning to the problem statement — the bot's purpose is guidance, not comprehensiveness.
- The same prompt produces meaningfully different outputs across multiple sessions — prompting is a skill that requires iteration, just like design.
- Custom GPTs and research AI tools (STORM) serve fundamentally different purposes: one extends a conversation, the other automates a literature review.
- Design thinking is not just a framework to study — it is a genuinely useful lens for building AI tools. Applying it to ThinkBot itself made the product better.
- The most valuable feedback in testing came not from what users said but what they hesitated to do — the chip suggestions became the most-used feature because users froze when faced with a blank input.
Without design thinking, the natural approach would have been to start building immediately — writing code before understanding what users actually need. The Empathize and Define phases reoriented the work around a specific user problem: students who freeze at the blank page of design thinking. This informed every subsequent decision — the welcome card, the suggestion chips, the stage navigation. The framework did not slow down the process; it prevented building the wrong thing.