Show HN: Revise – An AI Editor for Documents | AI Paper Digest

TL;DR Highlight

An AI-integrated word processor that lets you choose between OpenAI, Anthropic, and xAI models for document editing, correction, translation, and summarization — all in one interface. Tighter AI agent integration is what sets it apart from Google Docs/Word.

Who Should Read

Developers building AI document editing tools or similar services, or product engineers looking for ways to integrate LLMs into writing workflows.

Core Mechanics

Revise is a WYSIWYG editor using canvas-based rendering, built on a custom rendering layer rather than the browser's native text engine. This enables fine UI control but creates platform compatibility issues like missing macOS readline/emacs shortcuts (ctrl-a, ctrl-e, etc.).
You can choose between OpenAI (GPT-5.4 Mini/GPT-5.4/GPT-5.4 Pro), Anthropic (Claude Haiku 4.5/Sonnet 4.6/Opus 4.6), and xAI models directly. Users pick models based on task type and cost.
An AI agent sits in the editor sidebar, detecting inconsistencies in the document (timeline mismatches, conflicting numbers) and suggesting inline fixes. The goal is to eliminate the copy-paste-to-ChatGPT workflow.
A multimodal LLM feature converts PDFs to rich text. Even scanned or image-based PDFs can be turned into editable text.
User preferences are learned and remembered for personalization. Pronoun preferences (they/them), spelling conventions (US spelling), paragraph length, APA 7 format, Oxford comma, email max 120 words, etc. are stored and applied in subsequent interactions.
Word document and Google Docs import, PDF scan conversion, translation, summarization, paraphrasing, tone adjustment — all available in a single interface without tab switching, maintaining editing context while using AI features.
Built by a single developer over 10 months. The community viewed this positively as focused product development distinct from 'vibe-coded weekend project.' However, the subscription model (~$8/month estimated) and online-only requirement created barriers for some users.

Evidence

A developer interested in the canvas-based editor architecture also shared a detailed bug list: focus loss at 3-level tab indentation, double-click drag text selection not working, selection highlight offset, incomplete text restoration on undo, list style toggle collapsing hierarchy, table row vertical resize not available, etc.
A developer working on a similar project mentioned OnlyOffice + Python library (python-pptx, etc.) sandbox combo and Tiptap/PlateJS Notion-style libraries as alternatives. They mentioned considering open-sourcing their work too.
Multiple users expressed resistance to the subscription and online-only model. Reactions included 'I can already paste into Claude for free, why pay $8/month more' and shared experiences of formatting getting stripped when pasting into enterprise portals (Confluence, Wiki, Google Docs), making them stick with raw markdown.
Criticism that Notion already offers similar or more features, and that AI writing tools need 'deep editing' capabilities — logical flow, conclusion validity, redundancy elimination, essay structure rearrangement — to truly be useful. Current state was bluntly assessed as 'barely past the first step of ChatGPT copy-paste.'
Comments about MS Office and LibreOffice being too bloated while lacking lightweight alternatives. In this context, interest in Revise's editor quality was expressed even without AI features. Questions about team collaboration features (concurrent editing limits) went unanswered.

How to Apply

If building a custom canvas-based editor, use Revise's bug cases (platform shortcut compatibility, multi-level list toggles, selection area rendering offsets) as a QA checklist to catch issues early in development.
Planning an AI writing assistant SaaS? Consider starting with open-source Notion-style editor libraries like Tiptap or PlateJS as a base and just bolt on LLM APIs. The OnlyOffice + Python sandbox combo works when you need to handle Office formats like PowerPoint directly.
For building user preference personalization, Revise's approach of accumulating natural-language preference rules (pronouns, tone, format, etc.) in the system prompt works without any fine-tuning. Store per-user preference lists in DB and append them to the system prompt on every request.

Terminology

WYSIWYGWhat You See Is What You Get. An editor where what appears on screen matches the final output. MS Word and Google Docs are classic examples.

canvas-based renderingDrawing text and layouts directly on an HTML <canvas> tag instead of using the browser's default text input elements. Enables fine visual control but requires manually implementing OS-native text behaviors (autocomplete, accessibility, shortcuts).

multimodal LLMA language model capable of processing multiple input types beyond text, such as images and PDFs. Leverages image recognition to convert scanned PDFs to text.

AI agentAn AI system that goes beyond generating responses — it autonomously analyzes documents, executes edit suggestions, and performs continuous tasks.

rich textText containing formatting information like bold, italic, headings, and tables. The opposite of plain text, which has characters only.

system promptA hidden instruction given to an AI model before conversation begins, specifying roles and rules. Injecting personalization preferences here makes the model follow those rules in every response.

Show HN: Revise – An AI Editor for Documents