The UEFN AI setup

Your AI has great UEFN tools and no idea what they do

You ask it to make the room darker. It edits the point lights. Nothing changes. It tries again, lower. Then exposure, a post process device, the fog, the sky light.

Hours later you find it: a Day Sequence actor owned the sun, the sky and its own fog, and overrode every edit.

No error. No warning. Every single edit reported success.

That was my build. I wrote this pack so it would stop happening to me, then tested it until it held.

Get the UEFN setup
$500, paid upfront. You pick your 45 minute slot on the next screen.
A dark UEFN style stone level with small point lights on the walls that stay dim while a large glowing sky system above overrides them.

One Day Sequence actor owns the sun, the sky and the fog. Every wall lamp here is on.

Most platforms fail loudly. UEFN fails quietly.

Epic's MCP is one of the good ones. The problem is your agent has no model of the platform, and every quiet failure looks like success.

What your agent sees
What actually happened, and the skill that catches it
The property write returns true
A construction script threw it away on the next tick. uefn-debug
set_properties returns false, no message
Wrong argument shape. Nothing happened at all. uefn-mechanism
Fog density set to 0, then set back
The fog component unregistered itself. It does not come back. uefn-lighting
The fix is verified in the editor
Editor edits never reached the running game. There's a push gate in between. uefn-mcp-session
A screenshot proves the fog is broken
Device VFX never render in the editor. The screenshot can't contain the thing you're looking at. uefn-debug

Nine skills your agent reads before it touches anything:

uefnuefn-mechanismuefn-mcp-sessionuefn-lightinguefn-verseuefn-audiouefn-itemsuefn-interactionuefn-debug
About 1,000 tokens to answer a symptom. The fog misdiagnosis cost 640,000. That's the whole trade.

It was tested, which is the whole point

Anyone can write a document telling an AI how a platform works. Here is what happened when this one met real developers and a live editor.

640,000
tokens burned on one fog misdiagnosis, before the pack
Two agent runs reasoning from screenshots that could not show the fog. Measured on a production build.
141 of 141
runs loaded a skill on their own
On deliberately sloppy, lowercase prompts. No invoking.
15 of 47
scenarios passed round one. That's 32%
On a pack that read as finished. It knew the answers and failed to surface them. Then it got rewritten.
12
live editor tests
Ten written claims put in front of a real UEFN instance. Four were partly wrong. One was dangerous.
No amount of careful writing finds those. Only a live editor does. An untested knowledge pack is a confident knowledge pack, and confidence is exactly what's going wrong in your agent right now.

What you get, in hand

Six things, all of them yours to keep.

In your hands
1
A 45 minute call with me
What you're working on, what keeps breaking, and a look at your setup so I know exactly what to build.
2
The tested skills pack, installed and wired
All nine skills in your runtime, with the router loading on its own.
3
Skills built for your project
Your kits, your pipeline, your naming, the traps your team already learned the hard way. Same format, same tests.
4
Install videos
Short recordings of the exact setup, so your team can repeat it without me.
5
A follow up call if you get stuck
Another call with me, not a support ticket.
6
One year in the PlayUGC Skool
Where I and the community post skills, scripts and lessons for UEFN and Roblox.

Pay, pick your 45 minutes, I build it and send it over. If anything's off, we get on another call.

Flat fee, paid upfront. No equity, no revenue share, no retainer. You keep all of it: the pack, the skills built around your project, the videos.
Get the UEFN setup
Works with Claude Code, Codex, GPT, Grok or whatever your team runs.

Questions people actually ask

This isn't you yet
  • You haven't started in UEFN.
  • You want someone to build the game; this makes your agent better, it isn't dev for hire.
  • You don't use an AI agent at all. There's nothing here to install into.
Which AI tools does this work with?
Claude Code and Codex read the same open skills format out of one folder, so those two are a straight copy with a different destination. For GPT, Grok or anything else your team standardized on, we sort out the destination on the call. The content is plain text, so it travels.
What if it doesn't help?
Tell me and we get on another call and fix what's wrong. I won't hand you a number of hours saved. What I can point at is what got measured: 141 of 141 runs loaded a skill unprompted, and 32% of the pack failed its first round before it got rewritten. I'd rather show you the failures than a claim.
How fast do I get it after the call?
A real date on the call itself, once I've seen your setup. Installing the pack is the fast part. Building skills around your project is the part that takes time, and I'd rather give you an honest date than a quick one.

Give your agent a model of the platform

One 45 minute call, then a tested skills pack wired into your setup and new skills built around your actual project. Plus a year in the Skool.

$500, paid upfront. One time.
Get the UEFN setup
You pick your slot right after payment. No scoping thread, no proposal.