Stop Using Claude Code the Normal Way

Description: This video compares two implementations of the same content-creation app built with Claude Code: one via a standard single-session workflow and one using a long-running agent harness. The presenter highlights failures caused by context-window limits, then demonstrates a robust long-running approach that includes feature decomposition, regression testing, UI-driven verification, and Nano Banana thumbnail generation. The tutorial also covers local setup, practical options, and trade-offs between speed and test coverage.

– Demonstrates how single-context runs lose critical state and miss advanced features (editable prompts, card manipulation, thumbnail images).
– Explains a long-running harness: initializer agent, feature list stored in SQLite, coding agents that implement features and run regression tests.
– Shows a polished UI: app spec generation via Claude, history/revisions, editable system prompts, thumbnail editing/upscaling with Nano Banana.
– Provides setup and operational guidance: YOLO (fast) vs tested (slow, robust) modes, token/usage considerations, and practical tips.

Quotes:

Longunning agents are the future of Agentic Coding and Vibe Coding.

This is just stupidly impressive.

You can run this while you’re sleeping or over weekends.