I mass-accepted every AI code change for a week
Claude Code has an insanely powerful... and dangerous feature. It's called "dangerously allow browser." It auto-accepts every edit without asking permission. I turned it on because I was tired of clicking approve like one of those wooden drinking bird toys. Day 1: Fast. Exhilarating. Day 3: My codebase had 14 files I didn't recognise. Day 5: I found a function that called itself in an infinite loop. I didn't panic. But I did spend an entire evening untangling changes I'd never reviewed. Reverting commits. Re-reading diffs line by line. Doing manually what the tool was supposed to save me from. The lesson wasn't "don't use it." The lesson was: pair it with plan mode. Claude Code has another feature most people skip. When you enter plan mode, the agent stops coding and presents a markdown document of exactly what it's going to do. Every file it'll touch. Every function it'll change. The reasoning behind each step. You review the plan. Edit it if needed. Then send the agent to work. Now my workflow looks like this:
The combination is 10x faster than approving every edit manually. And 10x safer than blind auto-accept. Here's why this matters beyond Claude Code: Every AI tool has the same tension. Speed vs control. You either slow yourself down reviewing every suggestion, or you let the AI run wild and clean up the mess later. The answer is never "more speed" or "more control." It's sequencing them correctly. Plan first. Then execute fast. This pattern works for AI writing tools too. Don't let ChatGPT draft and edit simultaneously. First, agree on the outline. Then let it write. Review the structure before you review the prose. Most people use speed mode or control mode. The ones getting results use both — in the right order. Want to learn the exact prompts and workflow frameworks I use to get more from AI tools? Check out my Prompt Writing Studio course |