CleanFix Dev

THE TOOL LIST Video/audio: FFmpeg, DaVinci Resolve, WeftCut, FilmCraft, Audacity Images: GIMP, Inkscape, ImageMagick, PhotoCraft, VectorCraft, rembg 3D/render: Blender, Intel OIDN Cleanup: Real-ESRGAN, RIFE Explainers: ManimGX, Anatomy

The cheapest token is the one your agent never spends. I let Grok Bot drive tools I already own and run the heavy lifting on my own machine where I can, so the model mostly does the thinking. Setup below, point by point. 🧵

I run a ChatGPT dot and a Grok Bot side by side. What made both useful: giving each one its own lane. Not a ranking, just my setup, step by step. 🧵

Standouts 🌟 ComfyUI Character Card Studio (JGRFW on GitHub, v1.2.1, MIT): all 9 layout checks passed. No image generated. Update: Earshot by @lexnlin, 0.3.0 in Ed4, now 0.4.0. We rendered a whoosh, listened, and it's clean. Catch: one sound only; a wider test run stalled.

MOSS-Image-SFT by @Open_MOSS: open text-to-image model, Apache-2.0. On the official hosted demo, a prompt for a red mug with a label reading HELLO gave a 512×512 image in 12.2 s. Mug and lettering were right. Catch: one image, hosted only. We did not run it locally.

ComfyUI-DBS by @geneLab_999: splits audio into speech, music and effects stems. MIT. We ran its BSRNN and TIGER models on CPU on a 4-second test tone. Stems summed back with error under 6×10⁻⁸. 23 s and 43 s. Catch: no real soundtrack, no listening test, no ComfyUI run.

WhatAreYouBuilding.AI by @netadror: a hosted product directory with an API and MCP tools. Neta Dror is a woman builder, per her own site. Not open source. Listing, filters and leaderboard returned HTTP 200; MCP showed 9 tools. Catch: submitting and the full site flow untested.

Yara Reads by Eli Finer (finereli on GitHub): early reading practice where tapping a word splits it into sounds. Hosted app: 7 books loaded, "six" split into s/i/x and we heard it, pages turned. Catch: local Windows story checker failed: bad path. Teaching quality unrated.

FieldCraft by Tangency Labs: engineering simulation via CLI or an AI agent (MCP). MIT OR Apache-2.0. Windows v0.7.0, CPU: a sample cantilever passed its model check. A -1000 N load gave about +1000 N reaction, 0.24 mm max displacement. Catch: early software, GUI not tested.

Devs Helping Devs: Edition #8 🛠️ The Beyond-Code Toolkit We tested 5 tools plus 2 standouts: an engineering simulator, a kids' reading app, a builder directory, an audio splitter and an image model. Each tested at its own level, and we say which. #DevsHelpingDevs 👇

MARNE V2T just had its best week yet. Real breakthroughs, and I'm very close to the goal I set. Still building in the open, still not done, and I'm fired up.