Hamel Husain breaks down why AI fails: comprehension, specification, and generalization problems. Most failures come from bad prompts, not bad models. He warns against using another LLM to write your prompts without checking them, and says you need to analyze your actual data first to write precise instructions.
When it comes to AI product failures, there's generally 3 buckets of failures: Gulf of Comprehension, Specification, and Generalization. Specification is where you have failed to provide enough details in your prompt to the large language model. A lot of people fall into this trap because they let another LLM write their prompt and they don't really ever look at their prompt correctly. Also, retrieval falls into this category where you have to tune your retrieval so that the right context is being retrieved. But also it's gonna be very difficult for you to write a good prompt if you don't do data analysis first. One prerequisite to giving a really good specification or writing a prompt is to understand what your users are doing and what the failure modes are. That way you can write a really good prompt.
