Thoughts on using LLMs for coding, which can often be difficult and unintuitive, and why “vibe coding” is a fantastic way to explore the capabilities of LLMs
Online discussions about using Large Language Models to help write code inevitably produce comments from developers who's experiences have been disappointing.
Context & Ripple Effects
This sits in a practical-use debate where builders have documented prompting pitfalls in real-world LLM applications, while broader commentary argued models should be assessed by demonstrated capabilities rather than their failures. The article reframes disappointing coding results as partly a problem of learning an unfamiliar interface and feedback loop, not simply a binary judgment on model quality.
That framing foreshadows later coverage in which coding agents were described as useful, while retaining the caution implicit in lessons about common prompting pitfalls. It matters because experimentation can reveal where an LLM is helpful before it is trusted with consequential code.
First-order effects
- Developers are encouraged to use low-stakes exploratory coding sessions to learn a model’s strengths, failure modes, and interaction patterns before applying it to production work.
- “Vibe coding” shifts the immediate goal from reliable task completion to capability discovery, making disappointment more actionable as feedback about task framing and verification needs.
Second-order effects
- Teams adopting coding assistants will need workflows that preserve rapid experimentation while separating exploratory output from reviewed, maintainable code.
- Tool vendors face pressure to make model behavior easier to inspect and steer, since usefulness depends on how well developers can supply context and evaluate generated changes.
Third-order effects
- If exploratory use becomes a normal onboarding path, AI-assisted development may be organized around context preparation and human validation rather than one-shot code generation.
- The durable dividing line may be between tasks where iterative feedback is cheap and reversible and tasks where errors are costly; broad adoption will depend on tools narrowing that gap.
The trend: Coding AI is moving from a promise of autonomous code generation toward workflow-native assistance that developers learn, direct, and verify through iteration.