I made a tool that chains a small local model into a big coding model and auto-unloads VRAM between them
This covers a coding tool or code-capability update — useful for developers assessing workflow changes and reusable value.
PromptChain, a Streamlit app, now features significant updates. It chains a local "Prompter" model, which refines ideas into prompts, with a larger "Coder" model that generates cod…
A couple weeks ago I shared PromptChain here a small Streamlit app that chains two models: a little Prompter that rewrites your rough idea into a proper prompt, then a larger Coder that turns that prompt into code. The whole point is that on an 8–16 GB card you can usually only hold one model at a time, so it auto-unloads one before loading the other no manual swapping, no copy-pasting between two chat windows.
The comments last time turned into a real to-do list, so here's what's landed since:
- Reasoning models work properly now : blocks and DeepSeek-R1 / Qwen3 reasoning deltas stream into a separate collapsed panel instead of leaking into your prompt or code.
- Multi-file output : when the Coder emits several files, they render as per-file tabs with a zip download / save-all-to-folder.
- Pipeline profiles : save a whole setup (both backends, models, temps, system prompts) under a name and switch in one click.
- Persistent single-model chats : ChatGPT-style pages for just the Prompter or just the Coder; any drafted prompt jumps straight into the pipeline.
- Refine-in-place + version history : follow-up instructions ("make the board bigger") edit the code instead of regenerating, and every version is diffed and revertible.
The part I still like most: keep the Prompter local and point the Coder at a cloud model (OpenAI/Claude/Gemini). You fix the prompt for free on the local model, so the one paid generation lands right more often and you re-roll way less — frontier code quality without paying for every re-roll.
Local-first, MIT, no telemetry. Works with LM Studio, Ollama, or any OpenAI-compatible server
Genuinely after feedback both positive and negative. Also feel free to tell Prompter/Coder pairings that work well on your hardware