@KobinFlow: Addy Osmani, current core member of Anthropic's Claude Code team and former Google Chrome engineering director, just published a long post that directly calls out a common misconception in AI coding today: you're most likely "overprompting" Claude...

Addy Osmani, a current core member of Anthropic's Claude Code team and former Google Chrome engineering director, just published a long post that directly calls out a common misconception in AI coding today: you're most likely "overprompting" Claude. The most eye-opening fact: Anthropic officially cut Claude Code's system prompt by 80%, and the model's coding benchmark scores didn't drop at all! Many people are still writing prompts with the old "helicopter parent" habits they developed for models a year or two ago, piling on more and more words. Facing a new generation of models like Opus 5.5, the real experts are doing subtraction: 1. Don't teach it how to walk — give it just two things: one clear sentence stating **the ultimate goal** and **how to verify it was done right**. As long as you provide a unit test or validation script that outputs Pass/Fail, the model can grind through the loop on its own — no babysitting required. 2. Don't worship complex multi-agent architectures: Many people love to blindly stack subagents, but Addy is blunt: subagents don't make things smarter — they make latency and costs skyrocket. In practice, **a single agent with Skills** often beats multi-agent orchestration outright on both code quality and cost-effectiveness. 3. If you don't know how to do it, let the model interview you in reverse: When requirements are unclear, don't force yourself to write the perfect prompt. Let Claude ask you questions until it can compile a complete requirements spec. Then open a fresh session, hand it the spec, and let it start working immediately. 4. Don't be a human code scanner: Once the code is written, bring in a separate agent that hasn't seen the chat history to review the diff and run a round of fixes. Delegate all the low-level issues to the machine — humans should reserve their energy for making the final call on core business risks. **In one sentence:** Every time models take a generation leap, the first thing you should do isn't frantically adding rules — it's going into your prompt and **deleting code**. Give your prompts a diet every few months, and don't let the old habits of serving yesterday's models hold back what the new ones can do.