Why prompt refinement should run on your device

Key takeaways

• RefineIQ is a free, open-source Chrome extension that rewrites rough prompts in place — select text, click Refine, replace.
• By default the model runs entirely in your browser via WebGPU. There is no RefineIQ server; nothing you type can leave your machine, because there is nowhere for it to go.
• On-device AI is now genuinely good enough for narrow, well-scoped tasks — the trick is scoping the task, not scaling the model.
• The privacy claim isn't a policy, it's an architecture. Policies change; the absence of a backend doesn't.

The problem, honestly stated

Everyone who works with AI tools has the same tic: you write a rough prompt, get a mediocre answer, then spend three messages clarifying what you meant. The fix — rewriting the prompt to be specific before sending — is a skill, and even when you have it, it's friction.

Tools that "improve your prompt" exist. Nearly all of them do it by sending your draft to their server, which means your half-formed thoughts, internal project names, and pasted context now live in someone else's logs. For a tool whose entire job is to read what you're about to say to an AI, that's an uncomfortable amount of trust to ask for.

So we built the version we wanted to exist: the refinement happens on your machine, and the tool has no server to phone home to.

How it works

RefineIQ is a browser extension with one job. Select text in any input — ChatGPT, Claude, Gemini, or any ordinary text box — and a small Refine pill appears. One click and the rewrite lands where you were typing, with your original one Esc away.

Under the hood:

• A small language model in the browser. A one-time, consented download (~380 MB, with a lighter tier available), cached locally, accelerated by WebGPU with a WebAssembly fallback. After that it works fully offline.
• A narrow contract. The model does one thing: turn a rough draft into a single clearer prompt — preserving intent, never inventing facts, plain text, bounded length. It's not a chatbot; that narrowness is why a small model is enough.
• Voice input, also local. Speak a draft and an on-device speech model transcribes it; audio exists only in memory during transcription. We deliberately avoid the browser's built-in speech API because Chrome implements it server-side.
• Bring your own key, if you choose. Prefer a larger model? Connect your own provider key and prompts go directly from your browser to your provider. Still no middleman — there's no RefineIQ server to route through.

Why on-device is the point, not a gimmick

Three things become true when inference is local that no privacy policy can promise:

1. The trust surface collapses. "We don't store your prompts" is a claim you verify by audit. "There is no code path that could send them" is a claim you verify by reading the source — and RefineIQ is open source precisely so you can.
2. Sensitive contexts stop being edge cases. Legal drafts, unreleased product names, medical notes — the reason companies ban AI tools is the transmission, not the AI. A model that runs inside your browser sidesteps the ban's reason.
3. It works on a plane. Offline isn't a feature bullet; it's proof of the architecture. If it works in airplane mode, nothing is leaving.

The trade-off is honest: a ~380 MB model is not GPT-class. It doesn't need to be. Prompt refinement is a narrow, structured task — exactly the kind of work small local models are now good at. Scoping the task to the model is the whole design.

What this says about where we think AI is going

The default architecture of AI tools — every keystroke round-tripping through someone's API — is a phase, not a law. For narrow tasks, capable-enough models now fit in a browser tab. That inverts the privacy question from "can we trust them?" to "there's nothing to trust."

We build larger systems — agents, retrieval, orchestration — where server-side models earn their place. But the lesson from RefineIQ generalizes: put intelligence where the data already lives, and a whole class of risk simply doesn't exist. It's the same instinct as caging the agent inside the workflow — scope the intelligence tightly and the system around it stays trustworthy.

FAQ

Is it really free? Free and open source — no plans, no ads, no account. There's an optional support link in the popup; that's the entire business model.

Which sites does it work on? Purpose-built write-back for ChatGPT, Claude, Gemini, AI Studio, and Perplexity, plus any page with a normal text field. Per-site off-switch included.

What are the requirements? A Chromium browser on desktop; Chrome 113+ recommended for WebGPU. A Firefox build is maintained for portability.

Where do I get it? refineiq.obliviox.in — install, one model download, and the next prompt you send is a clearer one.

---

RefineIQ is built by oblivioX — engineering intelligent systems. If you're weighing on-device against server-side AI for your own product, the door is open.