Recap of Key Points

The biggest thing to remember about working with AI tools is that they are tools and you are in control. Once you give up that control is when the mess starts to grow.

To that end, just remember:

• Small steps, big wins. Think in terms of Git commits, the smaller, the better.
• LLMs are great at writing boilerplate, highly-predictable code. Basically stuff that's been written by other people routinely. It's not so good at interpreting business logic - it's your job to write that.
• Each Git commit should encompass a prompting session.
• Small steps allow for more exact prompts. Your prompt should be "closed", meaning "not open-ended or up for interpretation". When we created the data access layer, we had an exact SQL schema, and a pattern for Claude Sonnet to follow, and we got (mostly) what we asked for.
• Clear, concise instructions reduce hallucinations for the most part. Favor terse bullet points, and don't ask the LLM to interpret anything. "Favor best practices" is open to interpretation, whereas "always end a line of code with a semicolon" is not.
• Ask your LLM to review the code it created or, better yet, use a different LLM for variety. Have review guidelines ready to go.
• Take the time up front to 1) know what you want and what it looks like and 2) to write the instructions, spec, and other assets that will help the LLM give you the code you want.

You can do so much more with a logical code generator working for you. Does this mean you get to go home early? Yes. Does this mean you can do 3 times the work you normally do in your shortened day? Yes.

How you spend your time is up to you, which is what we'll talk about next.