Non-Coder Uses Gemini to Vibe-Code a 400-Line Modular Pipeline for LoRA Dataset Prep
ResponsibleKey1053 · reddit · 2026-08-19
A non-coder Reddit user shared how they vibe-coded a modular Python script suite with stock Gemini, unified under a ./run.sh, to manage messy image/video model data.
Key points:
- They ditched browser/Gradio tools due to crashes and missing multithreading; Gemini starts spilling out of bounds past 200 lines, capping usable output around 400 lines.
- One use was building a LoRA dataset to train a zoom LoRA — it trained, but underperformed because the base model already had that capacity and the prompting was simply bad.
- Another use: pulling and scaling reference clips for wan2.2, with a modular design allowing future add-ons.
- The author compares the vibe to early-2000s warez culture and invites others to share their weird niche tools.
More from coding & agent
- Dual 4090 local computer use setup: Qwen for reasoning, GUI-Owl for action — doodlestein · 2026-08-19
- html-video turns HTML/CSS into local MP4s with AI soundtracks — tom_doerr · 2026-08-19
- Opus 5 picks materials and generates poses for robot configurator — freelerobot · 2026-08-19
- Optimizing MCP: Cutting tool schema context from 54k to 1.3k tokens — Pitiful-Surround-285 · 2026-08-19
- Claude Code Skill Automates Home Assistant Configuration & Dashboards — tom_doerr · 2026-08-19
- Luthn: Open-source local agent memory layer with access control and auditing — Illustrious_Tell_741 · 2026-08-19