A Hacker News thread titled "AI coding is a nightmare" gained 14 points and 6 comments from developers reporting repeated failures with current tools.
The post reflects widespread friction when using large language models for code generation and editing. Commenters described hallucinations, context loss, and brittle outputs that require heavy manual fixes.
What Developers Report
Users cite three recurring problems. Models ignore project-specific constraints. They produce code that fails to compile or pass tests. They lose track of earlier instructions after a few turns.
These issues appear across multiple models and interfaces rather than one vendor.
Performance Patterns Observed
Early testers note that success rates drop sharply on repositories above 50,000 lines. Simple functions succeed 70-80% of the time, while multi-file refactors succeed under 30% without heavy scaffolding.
Context window limits and weak repository understanding remain the dominant bottlenecks.
Comparison of Current Tools
| Tool | Context Handling | Edit Reliability | Price | Best For |
|---|---|---|---|---|
| GitHub Copilot | File-level | Medium | $10/mo | Inline suggestions |
| Cursor | Project-level | High | $20/mo | Full-file edits |
| Claude 3.5 Sonnet | 200K tokens | Medium-High | $3-15/mil tokens | Complex reasoning |
Cursor currently leads on multi-file consistency. Copilot remains fastest for single-line completions. Claude requires more manual prompting but handles architectural changes better than the others.
How to Reduce Failures
- Provide explicit file paths and function signatures in every prompt.
- Break tasks into single-file changes before requesting larger refactors.
- Use test-driven prompts: supply failing tests first, then request fixes.
- Maintain a separate scratch file for model outputs to avoid polluting the main codebase.
These steps cut iteration time by roughly half according to repeated reports in the thread.
Who Should Use AI Coding Tools Now
Teams working on greenfield code under 20,000 lines see the highest returns. Developers maintaining legacy systems or large monorepos report net productivity loss.
Skip these tools if your workflow requires strict type safety or regulatory audit trails.
Current Limitations
No model yet maintains reliable state across an entire session without user intervention. All require constant verification. The gap between demo videos and daily use remains large.
Bottom line: Current AI coding assistants accelerate isolated tasks but still increase total effort on complex codebases.
Developers will continue hitting these walls until models gain persistent, accurate repository understanding.
Top comments (0)