Most AI coding tools are built around one promise: speed. Type a prompt, get code, move on.
But after a while I noticed the speed was costing me something. The AI would edit files I didn't expect, delete things I still needed, or run a command I hadn't read. I'd spend the next hour undoing work I never approved.
The problem with "just do it" AI
When an AI agent has full permission to write, delete, and execute, every action is a small act of trust you didn't explicitly give. For a throwaway script that's fine. For a real codebase it's not.
The failure mode isn't dramatic. It's quiet:
- A helper function gets "cleaned up" and something subtle breaks
- A config file gets rewritten with slightly different values
- A dependency gets bumped and nothing works the next morning
What I actually want
I want an assistant that treats my codebase like it belongs to someone else. That means:
- Ask before writing. Show me the change, let me approve it.
- Show diffs, not summaries. "I updated the auth logic" tells me nothing. A diff tells me everything.
- Checkpoint every step. If step 4 was wrong, I want to rewind to step 3 — not start over.
- Ask before running commands. A shell command is the most powerful thing an agent can do. It should be the most gated.
The trade-off is worth it
Yes, this is slower per action. But it's dramatically faster per outcome, because I'm not spending my afternoon fixing things I never agreed to.
If you're on Windows and want an AI editor that works this way, that's the idea behind Arcwright — it's free to start. But the principle applies to whatever tool you use: the AI should earn your trust one approved change at a time.
Top comments (0)