Descriptions:
David Gomes, an engineer at Cursor, delivers a conference talk explaining how the team replaced approximately 15,000 lines of code powering Cursor’s git worktree feature with a roughly 200-line markdown skill — a refactor shipped alongside Cursor 2.0 in October 2025. The talk is one of the clearest real-world examples of the “markdown as code” design pattern applied at production scale by the team that built the feature.
The original implementation required substantial bespoke infrastructure: worktree creation and lifecycle management, agent isolation and scoping, user-configurable setup scripts, a judging model that compared implementations via thumbs-up signals, harness modifications to keep agents on track, and cleanup tooling to prevent disk bloat. The replacement builds entirely on two existing Cursor primitives — agent skills and sub-agents — expressed as structured markdown prompts loaded on demand. A key operational benefit: because prompts are served from Cursor’s backend rather than bundled in the app client, Gomes can iterate on prompt quality without shipping a new Cursor version.
Gomes also details the “best event” command, where a parent agent spins up five parallel sub-agents across different models (Kimi, Grok, Composer, GPT, and Opus) on the same task, compares their outputs, and lets users cherry-pick and merge results. He is candid about what the refactor lost — not all functionality from the original system was fully replicated — making this a balanced rather than purely promotional case study in LLM-native software architecture.
📺 Source: AI Engineer · Published April 30, 2026
🏷️ Format: Deep Dive
