Developer suggests retyping LLM code to avoid debt

Software developer Ankur Sethi has proposed a deliberate workflow to combat the cognitive debt associated with AI-assisted programming, a practice he has refined since April. Writing in 2026, Sethi explains that while letting AI agents directly modify project files can theoretically make a developer 10x faster, it often results in a disorienting loss of control. By forcing the AI to output suggestions only in chat, developers can manually input the changes, maintaining a strong mental model of their codebase.

To enforce this boundary, Sethi configures his project agent files with strict instructions. The AI is forbidden from creating, editing, moving, renaming, or deleting any files, and it cannot run terminal commands or install dependencies. Furthermore, because Sethi is an experienced developer, the system is instructed not to explain basic syntax or programming concepts unless explicitly asked. This setup ensures the AI acts strictly as an advisory assistant rather than an autonomous agent.

For the individual practitioner, this approach reframes the relationship with AI tools. Sethi estimates that this manual process reduces his development speed from a potential tenfold increase down to a 2x acceleration. However, the slower pace forces developers to actively evaluate the code, making it much easier to catch hallucinations, refactor poor design choices, and clean up formatting on the fly. It mirrors the classic pedagogical advice given to novice programmers to type out code examples rather than copying and pasting them.

Ultimately, this method addresses a growing industry-wide risk where developers review massive, AI-generated pull requests without truly understanding them. Sethi warns that blindly accepting automated edits creates a systemic deficit of comprehension that the software industry will eventually have to pay back. By choosing comprehension over raw productivity, developers can ensure they remain the true authors of their software, preventing the accumulation of unmanageable technical and cognitive debt.