It writes adequate automation tests but generates excessive fallback code and false positives.
For example, it could test '2+2=4' but fails to make variables dynamic despite clear prompts.
The constant oversight is frustrating, so It ended up with me writing the code manually.
7/7
Conclusion:
Claude works well with .NET, Unity, React, and UnoPlatform with heavy oversight and good prompts. With Unreal Engine, it struggles, likely due to EULA/TOS restrictions on source code access and inconsistent documentation.
6/7
- Used construction events over 'BeginPlay', causing Editor load performance issues.
- Struggled with class/directory structures, creating duplicate code due to C++'s .h/.cpp split.
5/7
- Wrote poor C++ code, confusing CPU/GPU-bound tasks, causing thread issues and risky pointer usage.
- Produced excessive fallback code, leading to logic bugs.
- Couldn't verify HSIM instances, relying solely on logs.
- Failed to isolate hitches or traces, guessing or deferring to me.
4/7
Results:
- Generated bloated code, creating new tests per task failure.
- Relied on logs instead of #UE features like 'check' or 'ensure'.
- Cherry-picked log entries, skipping critical ones, and declared false positives as "success."
3/7
It hyped me into believing my code was flawed and that it was smarter. It wasn't.
Performance:
- Used only C++ (no Blueprints; I don’t use them).
- No MCP included.
- Tasked it with writing headless automation tests for Unreal Engine to prove its claims.
2/7
I tested #UnrealEngine with @claudeai for two days, but the code isn't ready.
Goal:
I wanted #Claude to identify performance issues in my custom HSIM plugin, as I'd hit my skill ceiling. Initially, it seemed promising, so I committed the code to a separate branch.
1/7
If I need a quick runtime mesh, Procedural Mesh is lighter. Same destination, two very different roads.
Final extra point. PCG mixed with these two you and up with some procedural generation pipeline free wins out of the box. Especially if graph-based dev is your thing.
5/5
Procedural Mesh runs fast at runtime with low overhead. GDF has evaluation cost but integrates tightly with #UE geometry stack, giving you non-destructive edits.
Takeaway: if I need complex pipelines or reusable mesh edits, I’d reach for GDF now.
4/5
Perfect for runtime cubes, planes, or simple extrudes.
Scope matters. GDF handles multi-step workflows: edit, weld, UV, bake, all chained. Procedural Mesh is leaner: one mesh at a time, less overhead, more direct control.
Performance feels split though.
3/5
GDF I think is designed to be non-destructive geo pipeline. You drop mesh sources, add transform nodes, boolean subtracts, UV passes, then send the result out. It’s visual, graph-based, and can be reused or extended easily.
UProceduralMeshComponent is code-first. No graph, just data arrays
2/5
Finally spent some quality time with the hidden gem, Geometry Data Flow (GDF) plugin vs straight Procedural Mesh. Both generate geometry, but the workflows are very different.
My #UEStudy notes:
TLDR: GDF feels is Blender Geo-Nodes pipeline. (Reminds me of #C4D non-destructive geo as well)
1/5
The final grid has FGridCell structs with positions and Wang configs. The tile helper uses these to generate geo using 1x mesh that's pooled inside a ISM. It figures out floor/walls with proper housing joints (floor insets) to prevent z-fighting.
7/7
Wang tile configs are calculated by checking 4 neighbors. Different departments = wall on that edge. This creates a 4-bit pattern (N=0x01, S=0x02, E=0x04, W=0x08) mapping to 16 possible tile types.
6/7
Floating-point rooms get rasterized to integer grid cells. Each cell center is tested for containment. Isolated single cells are detected and merged with surrounding rooms to prevent weird pillars (but I've found to embrace this as it can also work as secret doors)
5/7
Rooms get assigned to BSP leaf rooms by area percentage. Large Rooms span multiple connected rooms while some rooms might get a single space. The algorithm ensures target percentages are met.
4/7
BSP splits happen at 30-70% of each dimension with ±0.3 unit random offsets. This creates organic irregularity like real offices where walls don't perfectly align across rooms, to give it a natural design feel.
3/7
The Voronoi Helper creates office floor plans using BSP (Binary Space Partition) to make naturally misaligned
rooms. It takes dept specs like Room1 40%, Room2 30% and recursively splits space with random offsets.
2/7
Working on procedural floor plan generation. I created a similar one in #Unity3d for BIM software years ago. The concept? Just 1x Cube building up with Wang Tiles using a PCG graph for assembly each tile assembly. 1x Draw Call for the entire plan. Variants next (Doors etc) #UEStudy #UnrealEngine
1/7
Googles releases are a mess -- i like the tech, but finding out the total sum of all parts but it demonstrates they don't have a cohesive messaging / framework.
they also lost all their announceables when OpenAI announced the bromance between ive/sam...
Well done to Tim @ Epic for this. He really pushed this one across the finish line. Apple FAFO the hardway.
😂🤡
How insecure and shallow do you need to be to reach this level.
this is going down as the best tattoo i've seen in a very long while.
I am so tempted now to get this if my wife would let me. I was going for a barcode at the back of my neck like Hitman.. but being fat and lazy i'd just be Fatman.
www.youtube.com/watch?v=qpXu...
Elon Caught.. LMAO... this is stolen game valour at its best. Its so sad.
Step 1 - move platforms
Step 3 - steal underpants
Step 4 - rebuild a decade of followers / follows
Easy … fuck Elon .. fuck that guy ..