GPT's Lazy File Patching Problem - Partial Copies and Broken Imports That Waste Your Time
GPT's Lazy File Patching Problem
You ask GPT to modify a file. It rewrites part of it, copies some sections verbatim, skips others entirely, and leaves you with broken imports and missing functions. This isn't a rare edge case - it's the default behavior when files get longer than a few hundred lines.
What Happens
GPT's AUTO mode is supposed to pick the right model for each task. Complex reasoning gets the stronger model. Simple edits get the faster one. In theory, this is smart resource allocation. In practice, the file editing is where everything falls apart.
The pattern is always the same:
- You ask for a change to a specific function
- GPT rewrites the function correctly
- But it also "summarizes" the rest of the file - replacing actual code with comments like
// ... rest of the component - Imports that referenced the removed code break
- You spend more time fixing the patch than you saved
Why This Keeps Happening
Large language models generate text sequentially. When a file is long, the model takes shortcuts to stay within output limits. It prioritizes the section you asked about and treats the rest as context it can abbreviate. The result is a partial copy that looks complete at first glance but fails to compile.
How to Work Around It
Use diff-based editing tools. Instead of asking the model to rewrite entire files, use tools that apply targeted patches - replacing only the specific lines that need to change. Claude Code's edit tool works this way. So do most terminal-based coding agents.
Keep files small. If your components are under 200 lines, the model is less likely to truncate. This is good practice regardless of AI tooling.
Always verify. Run your build after every AI edit. Don't trust that the model preserved everything it should have. Check imports, check exports, check that nothing got silently dropped.
The lazy patching problem isn't going away soon. It's a fundamental tension between output token limits and file length. The best defense is using tools designed for surgical edits rather than full file rewrites.
Fazm is an open source macOS AI agent. Open source on GitHub.