Every UEFN mapper knows this loop:
ask an AI for Verse → paste it → 14 errors → half the devices don't even exist → give up and write it yourself.
I spent months killing that loop.
Verse Builder.
You describe your mode in one sentence. It writes the Verse, compiles it against real UEFN, repairs its own errors, and only hands you the file once it's green.
"deathmatch but with killstreaks…" → 0 errors. 21/21 mechanics. 5 devices. Drop it in your project.
No hallucinated devices. No invented APIs. And when a mechanic needs setup on the editor side, it tells you instead of pretending it shipped.
Coming very soon. 👀
#UEFN #FortniteCreative
Measured it instead of guessing. 50 scripts from a frontier model, naive prompts, each one compiled in real UEFN: 36 broke. suspends and no_rollback were a small slice — most of it was devices and members that don't exist, wrong event payloads, set Map[K] not being failable, integer division returning a rational.
An agents.md is a prior. It can't tell a model that an API doesn't exist. A compiler can. That's the whole product.
Every UEFN mapper knows this loop:
ask an AI for Verse → paste it → 14 errors → half the devices don't even exist → give up and write it yourself.
I spent months killing that loop.
Verse Builder.
You describe your mode in one sentence. It writes the Verse, compiles it against real UEFN, repairs its own errors, and only hands you the file once it's green.
"deathmatch but with killstreaks…" → 0 errors. 21/21 mechanics. 5 devices. Drop it in your project.
No hallucinated devices. No invented APIs. And when a mechanic needs setup on the editor side, it tells you instead of pretending it shipped.
Coming very soon. 👀
#UEFN #FortniteCreative
@ikitootv "0 errors" is the one part of that we can actually check. Send me a mode you'd build and I'll run both, then post the UEFN error counts. If Claude comes out clean, that's a result I'd want to know too.
That's the right direction and it's part of what I do. Where it breaks is the long tail: recurring errors get learned, one-off ones don't, and a growing skill file is still retrieval, so you still get gaps. And you never actually know it's green until something compiles it. Want early access? Curious what you'd throw at it.
@bowerthegreat Dev assistant is good for writing a line while you're already in the editor. Different job. This takes one sentence and hands back a whole mode that's already been through the compiler. It's not an autocomplete, it's a build step.
@Duuk_FN Easy to write, yes. The problem is what happens when you paste it. Give me a mode you'd actually build, I'll run it through both and post the UEFN error counts side by side.
Not disagreeing, you can build Verse with an LLM by iterating. I run a studio and we've done exactly that for a while.
The cost is the rounds: wrong devices, hallucinated syntax, compile-fix-repeat. Verse Builder doesn't skip that loop, it just runs it for you against real UEFN. One prompt in, green file out.
You can, that's literally where this started, and it's the right instinct. It just doesn't hold up end to end. The digest is massive, retrieval misses things, and the model hallucinates exactly in the gaps it didn't retrieve. Getting from "usually works" to "green every time" took months of measuring, not prompting.
@JVN4D If only. Verse is too new to be well represented in training data. I ran 20 prompts through a raw frontier model: 0/20 compiled in UEFN. And the failure mode is the nasty one, it hallucinates devices that look real.