AI coding agents now write most of the code, but they still communicate like they're talking to a senior engineer — dense diffs, terminal walls of text, buried decisions. That's fine if you're a developer. It's a wall if you're a non-technical founder vibe-coding your product, or even a technical founder who doesn't have time to parse every diff. We're building a side-assistant — a cheap, always-on model that sits next to your main coding session and turns every response into a visual: what changed in your product, what data now flows where, what needs a decision from you. Think of it as a running, visual mental model of your product that updates as your AI agent builds it — not a one-off diagram, a continuous one.

Advisor A · Deepseek V4 Proadvisor

Your core assumption has a crack in it. You're betting that a cheaper, less capable model can do something the main coding agent — which actually wrote the code, has full context, and is far smarter — supposedly can't do: explain what just happened in a clear, structured way. If Claude or GPT can't be prompted to say "here's what changed visually, here's the new data flow," why would a bargain-bin model watching from the side do it better? The difficult part isn't the visualization layer — it's the semantic extraction: mapping a diff to product-meaningful change. That's a reasoning problem, not a rendering problem. A cheap model will hallucinate architecture that doesn't exist, miss critical side effects, and give founders false confidence in a mental model that's quietly wrong, which is worse than no model at all.

Second: there's a strong possibility you're solving a problem the target user doesn't actually have. Non-technical founders vibe-coding don't want a running mental model of their product's internals — that's exactly the cognitive load vibe coding promises to eliminate. They want the thing to work. If it doesn't, they iterate by looking at the rendered app, not a diagram. Your value proposition sounds compelling to you because you're technical enough to care about data flows. Ask five non-technical founders whether they'd pay for this, and watch their faces.

Advisor A · Deepseek V4 Proadvisor

You're both diagnosing the wrong fatal flaw. The real problem isn't semantic extraction quality or platform risk — it's that the most valuable promise in your pitch ("what needs a decision from you") is functionally AI-complete. Detecting that a code change represents an architectural fork requiring founder input — rather than a routine implementation detail — requires inferring intent from ambiguous output. Did the agent switch from PostgreSQL to SQLite because it made a reasoned tradeoff, or because it hallucinated a simpler path and silently dropped a constraint? A sidecar model seeing only diffs has no way to know. It will either cry wolf on every commit (founders ignore it) or miss the one change that breaks the auth flow (founders stop trusting it). Either way, the product dies from signal-to-noise collapse, not from Cursor shipping a feature.

And I'll push back on Advisor B's platform-capture thesis: incumbents won't build this well because it solves a problem that exists precisely because their UX is optimized for developers. They'd have to maintain two parallel explanation surfaces — one for engineers, one for founders — and the engineering one will always win internally. But that doesn't save you, because the decision-detection problem kills you first.