<aside> 💡

Also relevant:

Understanding Code Assistance in Zed

</aside>

Large Language Models (LLMs) have revolutionized code and document generation, but using them for precise editing remains a significant challenge. While LLMs can generate impressive content from scratch, making targeted modifications to existing documents requires solving complex problems around edit representation, accurate placement, and verification. This challenge becomes particularly acute in professional environments where reliability and precision are paramount.

This document analyzes the current landscape of LLM-based editing systems, examining how different approaches tackle these challenges. Through a detailed review of systems like Tabby, Claude, Mandark, and others, we explore the fundamental tensions between precision and flexibility, automation and verification, and scalability and reliability. Our analysis aims to identify the most effective strategies for LLM-based editing and propose new directions for improvement, ultimately working toward more robust and efficient editing systems that can transform how we maintain and modify code and documents.

Overview

1. Prior Art Review

Key Systems:

• Tabby: Cursor-based edits (infilling, appending, completion) using structured prompts (<PRE>, <SUF>, <MID>).
• Claude Artifact Editing: Structured XML-like tags for targeted updates and rewrites.
• Mandark: JSON-based edits with line numbering and optional LLM verification.
• Claude-engineer: Simple search/replace mechanism for targeted edits.
• Void: Fill-In-the-Middle (FIM) prompts and diff-based edits for full file and targeted changes.
• Aider: Search/replace blocks and unified diffs for incremental and full file edits.

Key Insights: