AI Writing Tool

Overview

A browser-based writing assistant built on ProseMirror that ships zero client-side dependencies for AI features. The tool helps writers work through drafts by offering context-aware continuations, inline rewrites, and tone adjustments — all streamed in real time via a Cloudflare Workers edge backend.

The design goal was a tool that felt like a smart extension of your own thinking rather than something generating text for you. The AI features are surfaced subtly and never interrupt writing flow.

The Problem

Most AI writing tools treat the editor as a thin wrapper around a prompt box. You stop writing, switch modes, generate a block of text, and paste it back in. That context-switch is disruptive and makes the output feel foreign to your own draft.

I wanted to build something where AI assistance was woven into the editing experience itself — available without interrupting the flow state that good writing requires.

What I Built

• ProseMirror-based rich text editor with custom schema supporting headings, callouts, code blocks, and inline comments — built from scratch rather than using Tiptap or similar to maintain full control of the document model
• Inline AI completions — a subtle ghost-text suggestion appears after a short pause in typing, accepted with Tab, dismissed by continuing to type (similar to GitHub Copilot, adapted for prose)
• Selection-based rewrites — highlight any sentence or paragraph and choose from: shorten, expand, simplify, or change tone. The rewrite streams in-place character by character
• Context window management — the prompt sent to GPT-4 always includes the surrounding paragraphs and document outline, so suggestions are coherent with the broader piece, not just the current sentence
• Notion export — one-click export that maps the ProseMirror document model to Notion blocks via their API, preserving formatting

Technical Details

The most interesting technical decision was the streaming architecture. Cloudflare Workers support streaming responses natively, so GPT-4’s token stream is piped directly to the client without buffering. On the client, I wrote a ProseMirror plugin that progressively applies the streamed text as a series of transactions, so users see text appearing character by character in-place.

Doing this correctly required careful handling of the editor’s transaction history: mid-stream insertions need to be treated as a single undoable unit rather than dozens of individual keystrokes. I built a custom StreamTransaction wrapper that batches all stream events and commits them as one history entry when the stream closes.

SvelteKit was chosen over Next.js here because the editor component needed fine-grained reactivity for collaborative cursor positions, and Svelte’s compile-time reactivity model has zero overhead at runtime compared to React’s reconciler — relevant when ProseMirror is already doing significant work on every keystroke.