Ship Bigger Before It Breaks
Cursor just shipped subagents in 2.4 - multiple agents that explore your codebase in parallel before the main agent starts coding. My friend Ian who works at Cursor built a /council command that automates the whole thing. Today I'm testing whether this actually helps ship bigger features without the usual context rot. The real-world test: I have two separate apps right now. Ray Transcribes generates SRT files, and Vibestamps turns those into YouTube timestamps. I want to merge them into one. Can this new workflow handle a full feature migration on production code? This is basically the Fresh Chat Rule, but automated. Let's see if it works. LEARN TO SHIP Idea to Deployed Cohort (waitlist): https://startmy.ai Ship a real app in 3 weeks with auth, payments, and database. Next batch kicks off in February; 20 spots. GET THE TOOLS Ian's /council command: https://shaoruu.io/cursor/council https://cursor.com/docs/context/subagents Cursor: https://cursor.com Join our Discord: https://rfer.me/discord Key moments 00:00:00 Intro: Cursor 2.4's new skills and sub-agents 00:00:33 Ian's 'console' workflow for deploying agents 00:01:51 Today's goal: ship timestamp generation for Ray Transcribes 00:02:26 VibeStamps app and SEO benefits explained 00:03:54 Ray Transcribes speed and editing demo 00:05:52 Creating a new Git branch for the feature 00:07:20 Planning the VibeStamps integration in Cursor 00:10:24 Cursor's parallel agents kick off research 00:11:55 AI asks clarifying questions about requirements 00:13:36 Choosing a separate tab UI for timestamps 00:14:30 Q&A: Why I use Whisperflow for speech-to-text 00:17:31 Agent generates the initial implementation plan 00:18:08 My 'waterfall web development' philosophy with AI 00:20:23 Pivoting the plan to use Convex actions 00:28:03 Deep dive: How Ian's console skill spawns sub-agents 00:30:05 Installing the console skill into my Cursor config 00:31:51 Reviewing the updated architecture diagram 00:37:59 Using my 'phase plan' command for smaller steps 00:40:48 Kicking off Phase 1 with a new agent 00:44:59 Console skill re-adopts my 'dead code' finder 00:50:10 15 parallel agents find tons of dead code 00:54:38 Distracted by NVIDIA 5090 GPU scarcity 00:59:08 Phase 2: Zod schemas and Convex functions 01:04:02 Parallel agent handles user preferences 01:09:41 Insanely fast speaker diarization reprocessing 01:14:12 Q&A: Clawd.bot vs standard Claude Code 01:18:59 The 'sole document' that gives Clawd.bot identity 01:22:20 Feature testable locally; API key error 01:24:43 Timestamp generation works! 01:25:58 Live debugging an Edit button validation error 01:33:10 Convex real-time sync between UI and backend 01:38:35 Timestamp quality discrepancy discovered 01:39:51 15 agents investigate the prompt differences 01:43:40 Found it: missing critical prompt instructions 01:49:50 Back to Plan Mode for precise fixes 01:56:56 Shoutout to SpacetimeDB ("Convex for games") 02:01:20 New streaming schedule: Thu/Sat/Sun 02:07:39 Running local GLM on my NVIDIA DGX Spark 02:10:14 AJ's tutorial for Clawd.bot on AWS Free Tier 02:13:30 Reviewing agent-generated UI code 02:22:20 Testing the completed UI and backend 02:25:58 Agent fixes Convex validation bug 02:29:13 Full workflow demo: upload to timestamps 02:35:40 End-to-end test with auto-generated timestamps 02:38:43 Final 15-agent prompt investigation 02:43:40 All missing prompt logic identified 02:49:12 Final plan to hit 'gold standard' prompts 02:52:53 Distracted by the Zwift Tron bike challenge 02:56:56 SpacetimeDB chat about real-time game databases 03:00:30 Try Ray Transcribes (trial link) 03:01:20 Official new streaming schedule 03:02:30 Stream wrap-up CONNECT WITH RAY X (Twitter): https://x.com/RayFernando1337 #cursor #subagents #claude #opus #openai
- Jan 24, 2026
- 1.2K views
