Revise: AI Editor Built with Agentic Coding Tools and Y.js CRDT
Technical Implementation
Revise was built over 10 months using agentic coding tools, with the developer reporting never moving faster in their development career. The architecture includes a word processor engine and rendering layer built entirely from scratch. The only third-party library used is Y.js for the CRDT (Conflict-Free Replicated Data Type) stack.
AI Features
The editor integrates multiple AI models for document processing:
- GPT-5.4 Mini
- GPT-5.4
- GPT-5.4 Pro
- Claude Haiku 4.5
- Claude Sonnet 4.6
- Claude Opus 4.6
The AI agent works side-by-side with users in the same interface, providing inline proofreading and revision suggestions. It can detect inconsistencies in documents, such as timeline mismatches or contradictory statements about revenue changes.
Document Processing Capabilities
- Import Word documents and Google Docs
- Extract content from PDFs using multi-modal LLMs
- Proofread, revise, and perfect existing documents
- Export functionality with various settings
User Customization
Revise learns user preferences over time, including:
- Pronoun preferences (e.g., they/them)
- Feedback style preferences (direct vs. fluffy praise)
- Spelling preferences (US vs. other variants)
- Formatting preferences (APA 7, Oxford comma usage)
- Tone preferences (warm and precise, avoiding exclamation points)
- Accessibility considerations
- Document structure preferences (short paragraphs, concise drafts)
Technical Writing Preferences
The system supports specific technical documentation requirements:
- Performance claims must include benchmark setup, workload shape, sample size, and baseline comparisons
- Preference for concrete API contracts with request/response payloads, pagination rules, and idempotency semantics
- Breaking changes must be called out (API changes, data backfills, index migrations, client version gates)
- Requirements as numbered acceptance criteria with edge cases, retry behavior, and expiration details
- Trust in metrics like p95s, error budgets, queue depth, and incident counts over adjectives
📖 Read the full source: HN AI Agents
👀 See Also
Multi-operator Claude Code: Hub-based architecture for multi-agent sessions
A hub-based setup for Claude Code enables multiple people to attach to the same session, route subtasks across repos, and run headless agents in Docker containers.
The Human Creativity Benchmark: Separating Convergence from Divergence in AI Creative Evaluation
Contra Labs introduces the Human Creativity Benchmark (HCB), a framework that distinguishes objectively verifiable criteria (e.g., prompt adherence) from subjective taste (e.g., visual appeal) in evaluating generative AI for creative work. The benchmark reveals that no current model is reliably both correct and steerable, addressing mode collapse and the need for differentiated output.
Council: A Structured Dialogue Framework for Claude
Council — A Crucible is a structured dialogue framework that runs inside a single Claude context window, using persona framing to produce four distinct modes of engagement: rigorous interrogation, generative action, lived experience, and unformed intuition.
E2a: Open-Source Email Gateway for AI Agents with SPF/DKIM Verification and Webhook/WebSocket Delivery
E2a is an authenticated email gateway for AI agents that verifies SPF/DKIM on inbound mail, delivers via webhook or WebSocket, and supports outbound email with human-in-the-loop approval.