How our product shifted from trying to replace human PMs and designers toward showcasing the work that real people are already doing.
I see the development of the actual content itself to be a challenge. Accurate, meaningful, descriptive text has to be produced by someone, and maintained. This tool needs to be deeply integrated with that or a part of it. Probably difficult to be just a “jumping off point"... - Russell Sherman.
Following the trend
Six months ago our team at Splotch began exploring whether AI could represent user stories in a visual way. We’re product folks; could we make a tool to develop a fuzzy feature concept into a sprint-ready plan? Could AI do the legwork to create product flows for PMs and designers?
AI was reasonably good at mapping out flows based on text descriptions, but early feedback reminded us that the act of writing user stories and making high fidelity designs teaches you how the feature you’re building should actually work. In other words, without the process you don’t know if the output is any good.
By that point, we had discovered a novel way of translating text from an LLM into a visual model that could be edited and updated. But to what end? If we’re not giving PMs and designers a head start on their work, what problem were we solving?
Our first users speak out
We put together a group of sixty-odd testers across a variety of roles, including ops, logistics, sales, marketing, design, execs, consultants, system integrators. Some of these folks work for big tech—Meta, Google, Rivian, Shopify, DataBricks—and some worked for startups or consulting firms. We asked them: “What processes inside your company need to be mapped?” They said:
- We got processes! CUJs, SOPs, PRDs, BRDs, specs… all this stuff lives in our documentation. Except for a few CUJs and specs, it’s all text.
- If it’s important enough to need documentation, humans need to develop and check it.
- It would be nice to have visuals, but it’s a pain to create and maintain them. We only create them if the project is really important, to instigate conversation during design.
So we asked: what if visual aids came “free” as you develop processes? One of our testers, a design lead at Meta, responded with this:
“I can see people using Splotch to explain what they think, be confronted by what that actually looks like, then refine/change their mind. Currently our orgs huddle around spreadsheets and docs—perhaps this is a new way to gain alignment.”
😲 An aha moment
What if we were looking at our value add the wrong way around? What if, instead of trying to get the AI to do the work so humans don’t have to, we used AI to make the work humans are already doing more useful? What if our value is not in generating the work but in showing it off so other people will interact with it?
You may be asking the same question we did: isn’t AI already just a great big synthesizing, summarizing machine? Of course you can use AI to summarize any body of text. That’s how it works: it’s a summary of the Internet. The problem, as Ted Chiang points out in an excellent New Yorker piece, is that even though AI seems to give you a crisp, high-fidelity summary, it’s actually producing something more like a “blurry JPEG of the web.” It’s lossy, but you can’t tell.
But charts are different: you know at first glance that each node is a summary, a compressed version of the full story.
It turns out we’re not building an easier way to make flowcharts. We’re building a system to compress human work into visual form. We started with existing conventions—diagram types like flowcharts. Now we’re adding novel ways of compressing and uncompressing the trove of information that lives in documentation.
How should a human-readable compression system work?
For starters, we’re solving the blurry JPEG problem: in Splotch, every node is traceable—indexed back to its source material, so you can see the connections between summary and source. Full context is never lost.
We’re also making Splotch mutable and consistent, so that you can iterate on the compressed or uncompressed material and both sides will stay in sync.
And one more thing 😉 A tool that synthesizes and explains how things work needs to link into existing workflows and tools.
Splotch will be extensible!