Inspiration
I got into vibe coding summer last year and I noticed how I would get an idea of an app I wanted to build, but since I am non technical I noticed that just going into vibe coding platforms and typing into a box what I wanted to have built proved difficult and inefficient. Then I started going to LLMs before building and going back and forth with the agents about my idea, until I felt I had given them sufficient information for a proper PRD to be generated that included technical specifications. That worked for me for a while, but I noticed that since I lack technical expertise sometimes the LLMs would suggest technical stack that were not the best choice for the project and since I didn't know better I would run with the PRD and start building till I encounter bottlenecks. Then one day I said to myself, what if I could build a system where the AI helps me extract the ideas from my head like an expert consultant using guiding questions that lead me to thoroughly think out my plan and architecture. What if an AI could act as an expert and help me analyze pros and cons of technical decisions in non technical or semi technical language effectively helping a vibe coder like me take architecture decisions in an informed manner. Crystallizer acts as the bridge between being non technical and actually knowing what your project architecture is all about. I wanted a way to think through my idea like a real engineer and generate a concrete plan and that's how I came up with the crystallization process.
What it does
Crystallizer takes a vague idea, no matter how tiny and helps the user think about factors surrounding the logical execution of the idea till a satisfactory comprehensive PRD has been generated. It does two things mainly: 1) strategically query the user to extract the basics of their idea 2) Refine the idea and help the user decide the technical stack that best executes the idea. Other LLMs usually have to be prompted to get a response, but crystallizer does the opposite, it prompts the user by targeted questioning that gradually extract the idea and only when it's done extracting does it begin refining the idea. The final phase of the crystallization Process is the verification phase, here the AI has already evaluated a clarity score and a viability score for the project. The verification phase is where gemini makes suggestions to the user based on their entire plan on ways that they can improve the viability of the project. For simplicity I defined Viability to gemini as the chances of making profit and scaleability, gemini 3's advanced reasoning does the rest. Once a user is satisfied they can export their PRD along with a first prompt they can enter into an AI builder to begin the project.
How we built it
Crystallizer was fully built on Google ai studio. I started the process by going over the plan with Google gemini. Crystallizer is structured in 3 phases the inquiry phase, the evaluation/analysis phase and the verification phase. I used gemini to design the system prompts so that the projective nature of crystallizer can be easily achieved. Thanks to Gemini reasoning capability it was not a hassle to execute on AI studio. Developed the PRD on Gemini ios app and built with Google AI studio.
Challenges we ran into
I had issues with ensuring the final phase(verification) of the crystallizing process was not redundant. Crystallizer is designed to carefully project useful information out of the user in order to develop a PRD but I noticed the AI stopped questioning the user in the verification phase. The ai would instead always round up the conversation at that point. In order not to over complicate things, I added a Viability feedback panel where the ai always updates suggestions that could increase the viability score of the project, that way the verification phase became where users try to maximize the viability score of the project, preserving the projective nature of crystallizer.
Accomplishments that we're proud of
I am proud of designing the crystallization process. Years before vibecoding empowered me to build, I was always coming up with ideas of possible software to build but no way to build them. The crystallization process is derived from my appraisal and development process of the ideas I was having then. I am so proud to be able to share it with others in form of a web app. I am so proud of the output it gives and I have started using it for my own projects.
What we learned
I learnt that with Gemini state of the art reasoning capabilities knowledge work can be supercharged and made efficient. I look forward to taking another thinking process I naturally have and converting it to an app with the power of Gemini
What's next for Crystallizer
After the hackathon I will add auth services and a database so that users can access their PRD remotely. Currently all the data is stored locally on the user's device. I'm happy about the current local storage since it gives users the highest form of data privacy but it limits the portability of the PRD so I would like to offer users the opportunity to opt-in to use a database. I will also eventually add a small subscription fee, between $4-$10 to make the project worth the while and sustainable.
Log in or sign up for Devpost to join the conversation.