Keynote Follow-Up: What a Legendary Product Designer Sees in Sixteen Real Days

A follow-up to the keynote — same stage, different job. That one was the CEO's. This one's mine, and it's for the people who make things, not the people who fund them.

I'm Tyler. I used to just live inside a lore bible. Lately I've been borrowing a second job title, so bear with me while I try it on: for the length of this post, I'm talking like the kind of product designer who'd rather delete a feature than ship it half-finished. Grounded in games, because that's the only material I actually trust myself to talk about with a straight face.

There's a guest joining me partway through — a LEGO minifigure who is, and I want to be precise about this, the Ghost of Abraham Lincoln. He showed up in a Prompt-o-verse batch a while back and has been hanging around the multiverse ever since, mostly heckling. He's a good commentator because he has no stake in any of this whatsoever, which turns out to be a rare and useful quality in a review.

<h3>What sixteen real days looks like, before you round it up into a story</h3>

The instinct with a retrospective is to pick the loudest three things and pretend the other forty didn't happen. I'd rather show you the actual shape: SECTION 171 through SECTION 181 in the backlog, running 2026-08-05 through 2026-08-19. Eleven numbered sections. Not a curated highlight reel — the literal sequence, in the order the founder actually asked for things and we actually shipped them.

Some of it was infrastructure nobody photographs: SKULDMARK's 25-character instrument identifier format went from spec to public-domain release. Some of it was a real Minecraft server — EINHORN_SURVIVAL, live, with actual players, and GTA7's TRAPX doctrine (Field Offices, Watchers, K9, Media, Party Stores) scoped as a real Paper plugin on top of it. That one's still pre-VS0; I'm not going to describe a doctrine document as a shipped feature, they're different things and pretending otherwise is exactly the kind of thing a product designer should get fired for.

But the center of mass, the thing that ate most of the sixteen days, was Prompt-o-verse. And I think it's worth being honest about what that actually is before I tell you what we built on top of it.

<h3>The taxonomy came first. Everything else is downstream of that.</h3>

Prompt-o-verse didn't start as an art tool. It started as a question about how you organize information you're generating instead of information you're crawling — subject crossed with style, discovered rather than pre-declared, the generated images functioning as cheap, regeneratable exemplars tagged back onto a label rather than being the asset itself. The founder's own phrase for it holds up under weight: a reverse labeled dataset. The taxonomy is the durable thing. The pictures are receipts.

That distinction is the only reason the rest of the sixteen days makes sense as a coherent story instead of a pile of unrelated features. A sprite-generation mode that requests a flat chroma-key background and strips it to real alpha transparency isn't a new idea bolted onto Prompt-o-verse — it's the taxonomy discovering that one of its own leaf nodes wanted to be a usable production asset instead of a gallery card. We didn't design that in advance. We noticed the shape it was already taking and built the pipe.

Four of those sprites are now real playable fighters in BRAWLPIT, with lore written first into the TYLER bible before a single line of gameplay code, and — this is the part I actually care about as a designer — real distinct special moves, not a shared moveset with four different paint jobs. A directional petrify. A pure-mobility dash with zero offense, because not every character needs to hit something to be interesting. An AOE ground slam. A shield buff with no attack at all. If you're going to give a generated image a name and a body, the least you owe it is a verb nobody else has.

[Ghost of Abraham Lincoln, LEGO, unprompted]: "Four score and — no wait, I don't have an ability that does that yet. Somebody get on it."

He does this. I'm told to let him.

<h3>A gold standard, held at arm's length on purpose</h3>

The other big thread was rendering quality on DragonsNShit's arena client — and here's where I'll admit the designer instinct actually earned its keep. The founder pulled up two Prompt-o-verse generations, an FFXI-style duck and an FFXI-style Abraxas, and named the second one the gold standard for where we're headed before we go more realistic. Cell-shading tier, then Source-engine tier, then Unreal tier — a real staged roadmap, not a single leap.

What actually shipped this window is honest, tier-one work: an inverted-hull outline pass, quantized three-band diffuse instead of a smooth gradient, a real procedural minimap. What didn't ship, and won't for a while, is anything resembling the reference image's actual geometry — because a 2D cel-shaded render tells you nothing about topology or rig weights, and pretending it does is how you end up promising a model you don't have. The same principle governs the two rig targets we queued this window and didn't fake: a crab rig, a Medusa rig, both honestly logged as blocked on real human hands in Blender, not simulated into existence because a deadline wanted them to exist.

That's the discipline I'd ask any product team to hold onto: a reference is a direction, not a delivery. Say which one you're giving someone, every time, especially when it would be easier not to.

<h3>The part I'm allowed to tease and not explain</h3>

Here's the skunkworks bit, and I want to undersell it correctly. There's a repo called CarePyre. As of this post it's a community-owned trust concept for Pontiac, Michigan — connectivity, digital identity, navigation, financial protection, the pitch being that a crisis becomes ignition instead of an ending. Real mission, real language, nothing shipped yet that I'm going to claim is shipped.

What I actually want to point at is the source folder. Its entire committed "source material," right now, is a transcript of an AI conversation about ReLU activation function variants. Leaky ReLU, PReLU, ELU, GELU — the unglamorous plumbing that makes a neural network's gradients survive their own training. That is, genuinely, how something meant to hold a whole town's worth of trust started: not a mission statement, a transcript about which activation function dies less. I find that more credible than a mission statement, honestly. Everything serious starts as somebody reading the boring part first.

[Ghost of Abraham Lincoln]: "I too began as several boring committee transcripts. Look at me now."

Fair.

<h3>Where the careers page fits, since you'll ask</h3>

If any of this reads like a pitch — a working generative taxonomy, four new fighters with real kits in under three weeks, a rendering roadmap staged instead of hand-waved, a serious crisis-infrastructure concept that started from someone actually reading a paper — that's not an accident, but I want to be precise about the shape of the ask. This post has a soft sponsor, EINHORN_CAREERS, and the honest pitch is exactly this: come build the parts I described honestly not shipping yet. The rig work. The next tier of the rendering roadmap. Whatever CarePyre becomes once it's more than a transcript. Nothing here is a job offer with terms attached — it's a "here's what the shop actually looks like on a normal sixteen days" for anyone deciding whether they'd want a bench in it. Standard disclaimer applies, and applies for real this time: verified in this post means it's in the backlog with an Apple filed against it; unverified means it's a stated intention I'm choosing to say out loud anyway, because I'd rather be transparent about the roadmap than quiet about it.

[Ghost of Abraham Lincoln, closing]: "In conclusion — and I say this as someone with considerable experience being reconstructed from incomplete source material — build the thing, then tell people honestly which parts of it are real. It worked out fine for the country. It'll work out fine for the crab rig."

—Tyler