Landing page & market validation: Before writing a single line of code, I built a simple website and ran targeted LinkedIn ads aimed at product managers. A 2.5% conversion rate confirmed genuine interest and gave me confidence to go further.

PM interviews — round 1 (concept validation): Spoke with product managers to understand pain points and refine the positioning before committing to a full build.

Figma Design: Full design of component library and all main pages in Figma myself. 

MVP — knowledge base platform: Built the first version of Bonzaii using Lovable: a structured database and knowledge base for product teams. Entirely prompt-engineered — zero lines of code written by hand.

PM interviews — round 2 (the pivot insight): Feedback revealed the core problem wasn't storing data — it was knowing what to do with it. Teams weren't struggling to capture information; they were struggling to act on it. This changed everything.

Rebuilding as an AI co-pilot: Redesigned the product architecture entirely. Bonzaii evolved from a repository into an active co-pilot covering discovery, planning, delivery, feature management, notes & wiki, and an AI advisor layer.

Ongoing iteration with real users: The product continues to evolve with early-adopter PMs. Each feature is validated with users, tested with Playwright, and shipped through the same Claude → Lovable methodology.

No engineering background. No team. The core insight behind this methodology: the bottleneck in AI-assisted development isn't the code generator — it's the quality of thinking that goes into the prompt. I used Claude as my technical co-pilot and product critic. Every feature started as a structured conversation where Claude challenged my design decisions, flagged edge cases, and produced a sequentially numbered, granular prompt document ready to paste into Lovable.

One prompt. One task. If something broke, it mapped to exactly one thing to fix — making the process reliable, auditable, and scalable even working alone.

​

Architecture & prompt design in Claude

1. ​Discuss the feature, UX, data model, and edge cases. Claude challenges the design and outputs a precise numbered prompt document — nothing ambiguous, nothing skipped.

Prompt execution in Lovable

1. ​Paste into Lovable and execute. One task per prompt. Granular enough that a failure has a single root cause — debugging is fast, re-runs are surgical.

Test & validate

1. ​Playwright automated tests catch regressions. Manual visual checks confirm UX quality. Validated output feeds the next Claude conversation.

One of the most important design challenges was figuring out how to integrate AI in a way that genuinely adds value — without falling into the trap of bolting a chat interface onto a database and calling it AI-powered. I spent significant time researching and experimenting with different integration patterns: when should AI be proactive vs reactive? When does it earn a place in the main UI vs a side panel? How do you give it enough context to be useful without overwhelming the model?

The result is a layered approach — with the chatbot as just one deliberate component of a broader AI system:

Contextual AI actions

1. ​Suggestions and insights surfaced in-context — inside discovery, planning, and delivery modules — based on what the PM is currently working on. No context-switching required.

​Automated synthesis

1. ​AI analyses discovery data, groups insights, and surfaces emerging patterns — replacing hours of manual synthesis with a structured, reviewable output PMs can act on directly.

​Drafting & generation

1. ​From user stories to feature specs and roadmap narratives — AI handles the first draft, so PMs focus on judgment and decision-making rather than blank-page work.

​Natural language access (chatbot)

1. ​Yes, there is a chatbot — but it serves a specific role: giving PMs natural language access to their product database. Ask "what are our top pain points in onboarding?" and get a grounded answer from your own data.

The bottleneck is thinking, not coding. With the right prompting methodology, a non-technical PM can ship a production-grade product. The hard part is clarity of thought — breaking problems down, anticipating edge cases, writing specs precise enough to execute.

AI integration requires genuine UX research. Adding AI to a product isn't a feature decision, it's a design philosophy. The most valuable AI moments are the ones users don't have to ask for — contextual, timely, and grounded in real product data.

The "data → action" gap is the real PM pain. Every PM interview confirmed the same thing: teams don't lack information, they lack structured help deciding what to do with it. This is exactly where AI creates the most leverage.

Vibe coding works — but it needs discipline. Prompt-driven development is powerful, but requires the same rigour as traditional development: atomic tasks, clear acceptance criteria, automated tests, and an honest debugging process.