Why I Want My AI Coding Assistant to Ask Before It Acts

AI coding assistants promise speed, but their rapid action can come at a cost. These tools often edit files unexpectedly, delete crucial elements, or execute commands without explicit permission. The result is time spent undoing work that wasn't approved. While this may be acceptable for simple scripts, it's problematic for substantial codebases.

The issue isn't dramatic but insidious: a helper function might be "cleaned up," causing subtle breakdowns, or a config file could be rewritten with different values, rendering everything useless the following day. What I desire is an assistant that considers my codebase as belonging to someone else. This means asking before writing, showing the change for approval, and displaying detailed diffs instead of vague summaries.

When updating auth logic, a simple message like "I updated the auth logic" doesn't provide enough information. A detailed diff, on the other hand, reveals the changes comprehensively. Moreover, every step should have checkpoints. If step 4 goes wrong, it's preferable to rewind to step 3 rather than starting over. Granting permission for shell commands should be closely monitored, as they are the most powerful actions an AI can perform.

Although this approach may be slower per action, it's much faster per outcome, as I won't waste time fixing unapproved changes. For Windows users seeking an AI editor that operates this way, Arcwright offers a free solution. However, the principle can be applied to any tool: the AI should earn trust through approved changes, one at a time.