For larger projects I've found Linear + Obsidian + custom skills really helpful.
Linear for scoping work, acceptance criteria, and tracking tickets through PRs. I use independent review subagents to check that implementations meet requirements and actually use existing libraries/systems.
Obsidian for maintaining shared context (research notes, design decisions, rough plans that aren't ready for Linear, etc).
It's not perfect, and can feel pretty wasteful of tokens, but it's helped a lot with keeping agents aligned with my intent.
That said, I haven't actually been reading much code for the past month+, so there could be absolute garbage hiding everywhere lol. At this point, I've accepted that the next model release might be able to audit and clean things up faster than I could manually chase them down today π€·ββοΈ
Preview of a real-time @threejs destruction library I've been working on!
Lots more demos in progress, may do a technical breakdown on here of how it works.
Will open-source it once it's more polished.
Built on top of three-avbd: https://t.co/DG90OJEvUT
#threejs#webgpu
@The_MrAI@threejs Yep! Not sure if thatβs a criticism or just an observation haha. The goal was a quick test of AVBD for convincing real-time rope, sail and flag motion on an existing ship asset. Itβs definitely an approximation, but Iβm pretty happy with how it turned out so far!
@threejs Ship experiment with physics for ropes, sails, and flag simulated in real-time with three-avbd
https://t.co/DG90OJEvUT
Ship asset is from @polyhaven#threejs#webgpu
Been working on a @threejs WebGPU fluid simulator for a while. Lots still cooking, but here's a preview π§βπ³
More details and demos coming soon!
#threejs#webgpu
Ok the Teardown-inspired @threejs voxel destruction demo is live! is live!
Try it here https://t.co/bYmQEcXBQG added two more scenes - race track ποΈ and spacestation π
Code https://t.co/B9zCfBjIJ2
#threejs#webgpu
Good question! Honestly it depends a lot on the scene (how many pieces are touching, solver iterations, frame budget for rendering, what machine your on, etc).
But roughly, the solver alone handles ~100k+ boxes in a few ms per frame on an M4 Max. I'll stress test this scene tonight and post the results.
Repo + demos: https://t.co/0iiQwdftyc