Summary
Developer Web Dev Cody shares concrete lessons extracted from over 442 prompting sessions with Claude Fable 5.0 and 5.1, accumulated while building his open-source project Nebula 2i entirely through AI-assisted coding. The video starts from a viewer comment challenging the long-term value of vibe-coding and uses that as a jumping-off point for an honest, data-grounded analysis of what prompting patterns actually produce better results.
Key findings: effective bug-fix prompts consistently use the structure “when X happens, it should Y — here’s the bug,” with 77 of 442 sessions starting with the word “when.” Using ALL CAPS for hard constraints (“NEVER touch this file”) meaningfully improves agent compliance. For feature additions, language like “add in,” “extend,” or “enhance” signals that existing functionality should be preserved rather than replaced. The video also covers deliberately handing decisions to the agent when the implementation details don’t matter — a useful technique for reducing micro-management overhead.
The standout data point is a controlled A/B test comparing full Claude MD skill files against a completely bare context. The skills-enabled sessions used approximately 40% more tokens and were rated only “mildly better” by one judge model, with a second judge calling it a tie. Cody’s conclusion: for powerful frontier models like Fable 5.1, elaborate skill scaffolding often slows generation without proportionate quality gains — and the calculus only shifts when dropping to smaller, less capable models.
📺 Source: Web Dev Cody · Published September 15, 2026
🏷️ Format: Workflow Case Study
