AI coding is a nightmare. Am I the only one experiencing this?
A developer details frustrations with AI coding assistants, including the tendency to duplicate existing functions, bloat codebases with dead code, lack holistic awareness of the system, suffer from short context windows, and exhibit poor logical reasoning—often requiring major architectural changes for simple tweaks while producing code that fails to run.
Background
- The post is a complaint on Hacker News about AI coding assistants (GitHub Copilot, Cursor, Claude). These tools use large language models to generate and edit code from natural-language prompts.
- Key pain points stem from the "context window" — the amount of text an LLM can process at once (100K-200K tokens). When it fills up, the model forgets earlier context or degrades in quality.
- Common failures: AI adds duplicate functions because it only reads part of a file; it prefers adding code over deleting or modifying existing code, bloating the codebase; it treats each fix as an isolated task with no memory of the broader system.
- "Auto-compacting" refers to tools that summarize or drop early conversation history to stay within context limits, often losing crucial details.
- The frustration is widespread among professional devs: AI coding tools work well for isolated tasks but struggle with real, messy codebases that require maintaining existing logic.