Simon Griffiths

While we can all building applications at light-speed with AI, ensuring that the code produced follows best practice is more difficult. LLMs are fundamentally based on public code, and code that is public is not always based on best practice. Code that is truly scalable, fully secure and efficient is not always the default we see in public. Common example code is often simplified and designed for ease of use, not for performance.

A second challenge is that the LLMs are built on training data that is 6 months old or more, so new features and capabilities are often missed, and even if a recent feature is seen by the LLMs in training, true best practice will take time to evolve and be available in public.

Even worse, if you are an enterprise developer using an internal hosted LLM for privacy and commercial confidentiality, then it’s likely that your LLM will be older and the problem is exacerbated.

One way of addressing this is to have a set of agent SKILLS that encapsulate best practice. The skills can be updated with new features as they get released , and improved over time with best practice – thus breaking the dependency on LLM training data.

Kris Rice has published a great set of skills that can be used to address these problems, and it’s available on GitHub

This is designed as a base level of best practice, and the hierarchical structure of skills in the code generators allows you to extend with your own organisations best practice and standards with supplementary skills files, even down to the individual project level.

Take a look at the SKILLS and start using them to improve your generated apps.

https://github.com/krisrice/oracle-db-skills?tab=MIT-1-ov-file