Claude Code Creator: Delete Your Harness Every Six Months, Unleash the Model

In a YC interview published July 28, Boris Cherny, the creator of Claude Code, urged AI product makers to regularly delete their entire system prompt, skills, hooks, and harness code every six months, then add back lines one by one to test impact. This approach, based on ablation studies, reflects his empirical product iteration strategy: don't guess what the model needs; delete, test, and repeat. Cherny highlighted 'product overhang' — the gap where model capabilities outpace product integration — and 'unhobbling,' which means removing constraints to let models perform harder tasks independently. He shared an example where Opus 5, given access to OpenCV, could draw portraits without specific training. His unhobbling methods include: assigning harder tasks, running experiments, and enabling models to self-verify their work. He noted that Claude Code's system prompt was recently reduced by over 80% for new models like Opus 5. For AI users, he advised ignoring online gurus, taking an empirical approach, and treating models as colleagues rather than tools. He also recommended learning programming through application, not just theory.