A weekend, one LLM call, and zero features nobody asked for. That was the entire spec.
SoulCompass started with a two-sentence question. Could the Bhagavad Gita answer any modern human problem, and could an AI find the exact verse in real time. I gave myself a weekend to find out.
The trap with AI side projects is spending 80 percent of the time on the wrapper and 20 percent on the AI. I inverted that. The product had to do exactly one thing well: take a problem in plain language, return the shloka that speaks to it, plus a modern interpretation.
Next.js frontend. Gemini API for verse retrieval and interpretation. Web Audio API for ambient frequency sounds during reading. Vercel for deployment. That is it. No queue, no cache, no CDN gymnastics. Cold start latency was fine because the LLM call dominates the round trip anyway.
Prompt engineering. The whole product is worthless if Gemini returns a generic verse. I spent about six hours iterating on one system prompt with 30 problem inputs, checking that responses were contextually correct rather than surface-level pattern matches. This is where 80 percent of the perceived quality lives.
If you ship an AI product, the prompt is the product. Everything else is packaging.
One thing. I would have added a way for the response to be shareable as an image from day one. Every user who found something meaningful wanted to share it. I shipped the sharing on day four and lost a whole growth loop in between.
The lesson holds beyond this project. Ship the smallest version that reveals whether people care. Then ship the sharing before you ship anything else.
Claude Now Files Your ITR. This Is What That Actually Means.
AI filing income tax returns in India sounds like a novelty. It is not. It is a signal about where professional services are going and how fast.
Claude Design Is Solving Problems Nobody Thought to Give AI
Everyone expected AI to help with code and copy. Claude Design is now doing things in visual and spatial reasoning that most builders have not caught up to yet.