GPT-5 + Claude Code: Build a No-BS Sub-Agent With Radical Candor in Minutes
Detailed summary of the video
This short breakdown shows how to leverage the GPT-5 release and Cursor’s new terminal-based Agent to build a sharper, more reliable sub-agent inside Claude Code. The core idea is simple: many multi-agent setups get “helpful” and start hallucinating. You fix that by defining a strict “Principle Zero” for radical candor and truth, giving the agent a direct communication style and guardrails. The video walks through combining Cursor’s task-in-context workflow with a GPT-5 sub-agent in Claude Code, then layering on personality traits like INTJ and Enneagram Type 8 so it challenges assumptions, calls out weak logic, and prioritizes effective results over friendly fluff.
Chapters
00:00 Why this matters across platforms and why agent design is shifting
00:22 GPT-5 announcement context and Cursor’s new terminal Agent
00:33 Running tasks in context with Cursor Agent using GPT-5 by default
00:40 Creating a GPT-5 sub-agent inside Claude Code
01:11 Multi-agent pitfall: “helpful” agents give wrong answers
01:25 Plan: merge two designs to fix failure modes
01:41 Build the revised GPT-5 sub-agent and add Principle Zero: radical candor
02:00 Add personality traits (INTJ, Type 8) for direct, high-quality reasoning
02:19 Prompting toward efficient and, more importantly, effective outcomes
Credits:
Kieran Klaassen: https://x.com/kieranklaassen
Chris Royce: / christopher-royse-b624b596
Claude Code Subagents: https://github.com/NicholasSpisak/cla...