I spent a year teaching AI to read a mind map like a human argument — here's what changed when I stopped treating notes as text

Two days before my Product Hunt launch, I want to share the thing I learned the hard way — the thing that's not in any tutorial.

For about a year, I've been building Sensemaker, a tool that turns mind maps into coherent written narratives using AI. The pitch sounds simple. The execution was not.

The first version worked. That was the problem.

My initial approach was obvious: flatten the mind map into a list of nodes, concatenate them into a string, hand it to an LLM, ask it to "summarize."

It produced fluent prose. It also produced nonsense.

The AI had no idea that a node positioned far from the center carries less weight than one clustered with three siblings around a core idea. It didn't know that a floating, disconnected node is a tension — a thought that didn't fit anywhere, not a footnote. It couldn't read the argument embedded in the spatial layout. It only read the words.

So I rewrote it.

What I actually had to teach the AI

A mind map is not a document. It's a diagram of how a human was thinking at a specific moment. The spatial relationships carry meaning:

Distance from center = conceptual importance
Clustering = thematic grouping that the author felt but didn't name
Isolated nodes = unresolved tensions, open questions, things that didn't fit
Edge density = where the thinking is dense vs. where it's thin
Hierarchy depth = specificity — the further from root, the more granular

Once I stopped passing text and started passing structure — encoding the graph relationships, the spatial weights, the hierarchy levels — the output changed completely. The AI stopped summarizing and started reasoning. It could identify the core thesis the user was circling, name the tension between two clusters, and surface the buried assumption hiding in a leaf node three levels deep.

That shift took about four rewrites and a lot of maps that made no sense until suddenly they did.

What I learned about building AI features (that actually changed my approach)

The prompt is the least important part. Everyone optimizes the prompt. The real leverage is in how you structure the input. Garbage structure + great prompt = garbage output. Rich structure + decent prompt = surprisingly good output.
Users don't know what they want the AI to do — they know what they want to feel. Nobody told me "I want the AI to identify the spatial centroid of my argument." They said "I want to stop feeling like my ideas are a mess." That's a very different design brief.
"Good enough" AI output at launch will haunt you. I shipped a version that was technically functional but occasionally produced narratives that felt hollow — correct words, wrong insight. That version created more skepticism than the bug-filled version before it, because users engaged with it and then disengaged. A clear failure is easier to fix than a subtle miss.
RTL and multilingual support is not a feature — it's a statement. I built in Hebrew + English support (with proper RTL rendering) early. It made me think differently about the whole product: who is this actually for? Whose thinking style am I centering? That question improved everything.

Where I am now (48 hours before launch)

Sensemaker launches on Product Hunt on May 28th. The core loop:

Build a mind map (web, desktop, or local .mmap file)
Select any part of it
Get a 200–600 word narrative that identifies themes, surfaces tensions, draws conclusions

Free tier, no card required. For launch week, I'm offering 28 days of Pro access with code MAKE-SENSE (capped at 1,000 uses, valid through June 27).

I'm still learning what "done" means for a thinking tool. I suspect it means something different for every user.

If you're building something in the AI + knowledge-management space, I'd genuinely like to compare notes. What's the hardest AI behavior you've had to coax out of a model?

Demo (2 min): https://www.youtube.com/watch?v=J97Nb_ftYdA