Human-in-the-loop coding is badly needed!
Prompt-based approaches rely on instructing the AI (via system prompts or files like CLAUDE.md) to ask for help instead of guessing. In practice, developers have faced several issues with this method:
- Hallucinated Endpoints or Code
- Excessive Debugging Loops
- Ignored Stop Instructions
- Unwanted Output Artifacts
Prompt-only methods often lack a firm enforcement mechanism. The AI might continue generating code when it shouldn’t, leading to hallucinations and long debug sessions. These pain points highlight the need for a more reliable, tool-based solution.