A Hacker News thread titled "I'm going back to coding by hand" gained 38 points and 27 comments, with multiple developers describing their shift away from AI assistants.
The Trend Surfaced on Hacker News
The post and comments highlight a pattern: developers who once relied heavily on tools like GitHub Copilot and Cursor are disabling them for core work. The thread shows repeated mentions of context errors, hallucinated APIs, and time spent reviewing generated code.
Early comments note that AI suggestions often require more verification than writing the code directly.
How AI Coding Tools Operate Today
Current assistants predict tokens based on training data and the open file. They insert suggestions inline or in chat panels. When the surrounding context is incomplete or the task involves new libraries, accuracy drops sharply.
Users report that fixing these outputs frequently takes longer than typing the solution from scratch.
Benchmarks and Real-World Numbers
No public benchmark captures the exact time cost of verification, but community reports in the thread cite consistent patterns:
- 30-50% of suggestions require edits or rejection
- Average review time per suggestion: 8-15 seconds
- Net productivity gain disappears on unfamiliar codebases
| Approach | Typical Speed | Error Rate | Context Needed |
|---|---|---|---|
| Hand coding | 20-30 lines/h | Low | Full understanding |
| Copilot | 40-60 lines/h | 30-50% | Strong file context |
| Cursor | 50-70 lines/h | 25-45% | Project indexing |
Pros and Cons of Each Method
Hand coding:
- Forces deeper understanding of the codebase
- Eliminates review overhead for simple tasks
- Slower for boilerplate and repetitive patterns
AI assistants:
- Accelerate known patterns and standard libraries
- Introduce subtle bugs in edge cases
- Require constant context management
Alternatives and Direct Comparisons
GitHub Copilot, Cursor, and Continue.dev represent the main options. Copilot focuses on inline completions. Cursor adds chat-driven refactoring. Continue.dev offers open-source local models.
The HN comments indicate that developers who switched back to manual coding did so after testing all three and finding the verification cost too high for their specific workflows.
Who Should Code by Hand
Teams working on novel algorithms, security-critical systems, or small codebases benefit most from manual coding. Developers maintaining large legacy systems or generating tests see clearer gains from assistants.
Skip hand coding only if your daily tasks stay within well-documented libraries and patterns.
Bottom Line Verdict
The thread shows a measurable subset of developers achieving higher effective output by disabling AI tools when context quality is low. The decision hinges on task type rather than blanket adoption.
The shift reflects maturing evaluation of where these tools actually save time versus where they add hidden costs.
Top comments (0)