Show HN: VibeFarm – Modular composition layer for language models (app.vibefarm.ai)

Hi HN, I built VibeFarm to move beyond one-off prompting into modular composition. Instead of writing full prompts from scratch, you stack elements inside VibeCards that break language into structured controls you can lock, randomize, and remix. Think Ableton for prompts: lock what works, randomize the rest, and reuse the stack as a system.

Or imagine a prompt marketplace where, instead of sharing a single prompt, users exchange entire systems they can load, remix, and reuse.

What’s different

Structured slots: Subject, Context, Style, Goal for consistent, editable outputs

User-created cards: Generate new VibeCards or Elements directly from text. Everything in the system is user-extensible

Lock + dice: Freeze any card and explore controlled variation

Portable units: Save stacks as plain-text .vibe files for use across models

Combinatoric exploration: Even small stacks create practically infinite variation without drift

Semantic layers: Internal separation of linguistic, metaphorical, and analogical signals keeps style portable as models evolve

Language, not syntax: When everyone types the same phrasing, models converge on the same style. VibeFarm modularizes language itself, letting you explore tone, rhythm, and structure as creative variables.

Why it matters While most tools focus on interfaces or automation, almost no one is addressing the core material of LLMs: language itself. Ignoring that is like building Photoshop without touching pixels.

VibeFarm acts as a composition layer on top of models like Midjourney and Sora, helping creators break out of prompt grooves and preserve distinct intent across text, image, and video generation. It scales naturally as models evolve.

Try it Live demo (instant, no signup): https://app.vibefarm.ai

More info: https://vibefarm.ai

Under the hood React + TypeScript, Zustand, Node/Express, Neon Postgres + Drizzle, OpenAI API. The .vibe format is an evolving, minimal, versioned text spec.

All feedback is welcome. Happy to discuss implementation details, design tradeoffs, and next steps in the thread.