How to get the most out of your AI model

How do you prevent your AI assistant wasting tokens by generating something you'll probably never use? Much like when you're sat in front of a computer deciding what to concentrate your time on, AI assistants need to be told up front what they should be focusing on. Left to their own devices, they will happily generate pages of confident-sounding code and text that simply don't fit your project.

After having launched my own blog and an equipment management web app as practice projects, I'll tell you how to get the most out of your AI assistant.

It all boils down to defining exactly what you want the AI assistant to do. By doing this, the AI is laser focused on the task at hand and you'll no longer have to worry about token burn rate. But what would this look like in practice?

For me, it looks like a set of spec files. Before I let an AI agent anywhere near the code, I write down exactly what the feature should do, how it should behave, and what success looks like. That little bit of upfront effort saves a surprising amount of wasted output later.

I have posted a set of spec files for a MERN-based project on my GitHub account (link can be found above next to my avatar picture). Feel free to edit them for your own projects, and if you would like to, post a link to an update in the comments section below.

Bearing in mind I used this format before I heard of agent skills. The downside is you'll have to change the spec files for each individual project. That is fine when you are working on a handful of things, but it gets repetitive if you juggle many projects.

I've also gotten into the habit of getting the AI model I'm working with to generate a set of spec files in case I wish to regenerate the source code with certain changes (for example if a client decides to go with an alternative CSS framework). I usually include this with the project so that I don't lose context. It means I can rebuild the whole thing from scratch weeks later without having to re-explain everything.

If you're looking for something more generic that can work across multiple projects, then this GitHub repo might be something to test (not created by me):

https://github.com/addyosmani/agent-skills

The key takeaway is simple: give your AI a clear brief and it will reward you with focused, useful work. Skip that step and you will spend far more time (and tokens) cleaning up the mess.