where you fully give in to the vibes, embrace exponentials, and forget that code even exists.” — Andrej Karpathy Former OpenAI Founder and Tesla AI Leader, Founder of Eureka Lab
Coding Fig 1, 2: X Posts by Andrej Karpathy, 2023/1/24 & 2025/2/2 User Interest Natural Language Code Generation Working Software Enabled through visual interfaces, e.g. Claude, Lovable
on the sidebar by half” because I’m too lazy to find it. I “Accept All” always, I don’t read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it… Sometimes the LLMs can’t fix a bug so I just work around it or ask for random changes until it goes away.” — Andrej Karpathy Former OpenAI Founder and Tesla AI Leader, Founder of Eureka Lab
Coding Specification Implementation Debugging Traditional Coding Workflow: Code Generation Inspection Regeneration and Iteration Vibe Coding Workflow: ** done through natural language ** done through code
different tools to vibe code apps - 20+ Use Cases Collected - 4 Platforms Used Secondary Research with Literature Review: Read blogs and papers analyse popular case studies and workflows in vibe coding communities GenAI Assistance in Research • Critique and tighten my research question • Assisting with qualitative data analysis
• Does not know web/software development • Avoids file structures and manual debugging • Relies on conversational iteration Individual Creative Workflow • Works independently • No interest in GitHub collaboration workflows • Treats software like a personal creative medium Low-Stake Small- Scale Projects • Small-scale or disposable projects • Operates within token limits • Minimal concern for security complexity
Quantitative Evaluation Metrics: Effort invested Ease of use Smoothness of iteration Qualitative Data: Prompt Design Debugging Strategies Comparative analysis across different vibe coding platforms.
Simulation Productivity Meeting to Decision Transcript Summary Lifestyle Protein Intake Tracker Creative Music Visualisation Mapping the types of projects we explored through conversational software creation
What problem are we trying to solve? Creation How do we build effectively through Prompting? Improvement How do we identify meaningful improvements? How do we solve it through prompting? , , , Adoption Do we continue using this project? Vibe Coding can be divided into four stages.
What problem are we trying to solve? Creation How do we build effectively through Prompting? Improvement How do we identify meaningful improvements? How do we solve it through prompting? , , , Adoption Do we continue using this project?
Three Patterns of Vibecoding Ideation We found inspiration from online blogs. We want apps to fit our habits. We want to test the technical boundaries of Vibe Coding. Deeksha’s Polaroid App Coco’s Dictionary App Melissa’s Protein Intake Tracker App Deeksha’s Outreach Personalizer App Emery’s Oura Ring Biking Tracker App Coco’s Doom Scrolling Intervention App Replication Personalization Technical Exploration
Replit appear strongest overall. Claude has the most use cases due to familiarity with the platform. Effort scores for individual projects become increasingly less correlated to performance ratings. The various patterns show a large spread within a single platform. Performance depended heavily on the specific project context.” Fig 1: Comparison between our average evaluation scores on Lovable, Replit, Google AI Studio, Claude, and Base 44
Case Travel Planner on Claude and Lovable Low Technical Complexity → Similar Performance Fig 2: Emery’s Use Case Oura Ring Bike Tracker on Claude, Replit, and Lovable API Integration → Diverging Performance
What problem are we trying to solve? Creation How do we build effectively through Prompting? Improvement How do we identify meaningful improvements? How do we solve it through prompting? , , , Adoption Do we continue using this project?
Prompts Before Generating Code • Initial App Design • Structured Specification • Code Architecture How to Adopt in Vibe Coding? Instead of asking AI directly for an output, you ask AI to first design the reasoning structure for the task. What is Meta Prompting?
Prompts Before Generating Code • Initial App Design • Structured Specification • Code Architecture How to Adopt Metaprompting in Vibe Coding? Instead of asking AI directly for an output, you ask AI to first design the reasoning structure for the task. What is Meta Prompting? User LLM Vibe Coding Tool Product Take an Extra Step Here!
Structure Costly Metaprompting Solving already localised, specific problems High-Leverage Metaprompting Formalising ideas and inspirations into effective frameworks Structuring Ambiguous Ideas Translating Creative Ideas into Implementable Features Planning Systems Architecture Before Generation
Structure Structuring Ambiguous Ideas Translating Creative Ideas into Implementable Features Planning Systems Architecture Before Generation “I want to make a website that updates me on local screenings and concerts.” Initial Idea: Metaprompting: Break down with: • UI style • user workflow • data sources • APIs • architecture • scope boundaries
Poem Website Use Case “I have a poem I want to present in dynamic and interactive website. give me different ideas for my website based on the content of the poem, you need to include every line and the title of the poem but have creative freedom in your design. I would like you to develop prompts that I can give to vibecoding tools for building the website:” Fig 1: One of the prompts generated by ChatGPT 5.5 Instant (New Model Released) The prompt details technical and design requirements according to my initial prompt. • Design Requirement • Interaction • Style • Output
Structure Structuring Ambiguous Ideas Translating Creative Ideas into Implementable Features Planning Systems Architecture Before Generation “I want the website to feel more interactive.” Initial Idea: Metaprompting: ”Sections should animate sequentially on scroll, with toggleable visual elements and hover feedback.”
Structure Structuring Ambiguous Ideas Translating Creative Ideas into Implementable Features Planning Systems Architecture Before Generation “Help me design the architecture for a personal finance dashboard before generating code.” Prompt Input: Metaprompting: Define: • page hierarchy • component structure • state management • data flow • file organization
What problem are we trying to solve? Creation How do we build effectively through Prompting? Improvement How do we identify meaningful improvements? How do we solve it through prompting? , , , Adoption Do we continue using this project?
with words. After all, you have to know what you want to create; be able to judge whether the results are good or bad; and give appropriate feedback. This underscores how vibe coding isn't about eliminating expertise but redistributing it— from writing every line of code to knowing enough about systems to guide, troubleshoot, and evaluate.” — Ethan Mollick Wharton Professor
with the Product, Especially Testing the Edge Cases If you’re creating a Polaroid filter app, can it take HEIC files? (photos took on iPhones) Restart or Switch Tools Restarting a project only requires a new prompt with vibe coding and can bring about fresh results. Technical insights might be helpful to identify the causes of these bugs. Include Command in Prompts “Debug as you go along” Debugging with Positive or Negative Examples “Create a community feature on my food rating app like Beli” The minimum viable knowledge for effective debugging: your understanding of your needs and the tool you’re using.
What problem are we trying to solve? Creation How do we build effectively through Prompting? Improvement How do we identify meaningful improvements? How do we solve it through prompting? , , , Adoption Do we continue using this project?
Friction Users already have entrenched digital habits Technological Friction Vibe-coded tools often lack mobile-native convenience and ecosystem integration
problems became an important nontechnical skill. Creative Integration between Tech and Personal Life / Hobbies We can all be micro-entrepreneurs. Adoption might take off more with mobile integration. Vibe coding converges with our day-to-day habits. Flatter Learning Curve Software development usually requires a steep learning curve. Build Custom Tools for ourselves
three patterns of vibecoding ideation. Tool differences narrowed as prompting skills improved Creation: Meta-prompting is the most effective prompting strategy. Adoption is difficult because it involves habit change. Improvement: Nontechnical users can do debugging with limitations.
Emery, and Nadine Resources Individuals Blogs: • Andrej Karpathy’s Tweets. • Simon Wilison’s Web Blog. URL: https://simonwillison.net/ • Zara Zhang’s Blogs. URL: https://zarazhang.com/writing-profiles/ Forums and Discussions • Claude Artifacts for providing inspiration and tools for vibe coding. URL: https://madewithclaude.com/