Code generation

The LLM task of writing or completing source code from natural-language description or existing code context — the core capability behind GitHub Copilot, Cursor, and Claude Code.

Code generation is the task of producing source code from natural-language instructions, partial code, or both. The output ranges from completing the next few characters ("the rest of this for-loop") to writing entire functions, refactoring across files, debugging, or generating full applications from a description. It matters because coding is one of the use cases where LLMs have most clearly transformed productivity. Tools like GitHub Copilot, Cursor, Claude Code, Aider, Continue, and Windsurf are now standard parts of professional developer workflow. Estimates from Microsoft and others suggest these tools speed up everyday coding by 20-50% on appropriate tasks. A concrete example: type "// fetch user with retry and exponential backoff" in your editor and Copilot suggests a complete TypeScript function. Or describe a bug to Claude and paste your stack trace — it identifies the off-by-one error and produces a fix. Or paste a feature request and your codebase: Claude Code edits multiple files in coordination. Quality ranges enormously by model and task. Simple completions and well-defined functions are usually correct; complex multi-file refactors, novel algorithms, or anything outside the model's training distribution can produce subtly wrong code. Always test. Related: HumanEval, SWE-bench, agent, tool use, Cursor.