Summary
Theo (t3.gg) shares a month’s worth of hands-on lessons learned using Anthropic’s Claude Fable 5.1 for day-to-day coding work, distilling insights from Anthropic’s official prompting guide alongside his own agentic workflow discoveries. He walks through practical techniques for improving results, including how to structure agent commands, give models the tools and context they need to work autonomously for longer stretches, and even use OpenAI’s Codex to help refine Claude’s outputs.
The video includes real examples pulled from his own codebases, such as directing Claude to investigate a broken image-rendering bug across multiple failed pull requests, evaluate a competing contributor’s alternative fix, and decide autonomously which approach to merge — illustrating a shift toward giving models more decision-making latitude rather than choosing between options manually. He also demonstrates integrating Sentry’s MCP server into an agentic workflow to trace token costs, errors, and full request timelines when debugging agent-built applications.
Aimed at developers already using or considering Claude for coding, the video offers concrete, tested workflow patterns—including cautionary notes about “YOLO” autonomous merge workflows—rather than abstract theorizing about AI coding agents’ capabilities.
📺 Source: Theo – t3․gg · Published September 22, 2026
🏷️ Format: Workflow Case Study
