Running AI..

AI is finally becoming useful... to me. When it was just a "chat" bot, I couldn't have cared less. And when it could spit out a python script that I had to copy and paste into an editor and test and debug myself... I still wasn't impressed. But now that AI can "use tools", it is now becoming useful to me.

I am not an AI pundit or an expert by any stretch of the imagination. I have not drunk the cool-aide and think AI will change everything. I'm just some guy on the Internet and I do think things will change. So take these tips for what their worth.

You do not need to follow or even understand everything here before starting with AI. You most likely will not. As you gain more experience revisit this page and what didn't make sense at first, may start to make more sense.

Here are some of my "discoveries" when running AI.

First things to always remember

Keep these items in mind when using AI.

1. YOU ARE RESPONSIBLE!!! And you will be held responsible. If you ask an AI to do a thing and it does it. Or you ask AI to do a thing and it does something totally different. You are responsible for any damage or outcome that happens. Never forget that.
2. AI is like an idiot savant or a stochastic parrot. What AI can do is amazing, but AI is just a tool. It doesn't "know" anything. It doesn't "understand" anything.
3. AI uses human language, so it can seem human. It is NOT. AI is, at its core, a large (really large) multidimensional array. The fact that you can put words in and get words out something that make sense and can do things is like magic... but its just technology. Remember what Arthur C Clark said "any sufficiently advanced technology is indistinguishable from magic". Right now, AI is like that.
4. Hardware prices are crazy for running models yourself (Oct '26). Start with a monthly subscriptions to one of the "frontier" companies if you have never used AI. Once you get some of the basics down, you'll probably want try a different harness than the "frontier" company harness. Then you will start with a site like ollama.com, huggingface.co, nousresearch.com, openrouter.ai, etc and be able to select from a bunch of different models.

General usage guides

1. Harness matters a lot more than the model. A good harness with a marginal model can give you good results. A bad harness with a great model can also give you good results. The difference is the great model probably costs 3x more than the marginal model. Result matter.
2. Don't get hung up on the 'latest and greatest' model released. Find a model that can do what you want for the price you want to pay.
3. You need to have an idea / goal first. Take as much time as you need to refine that goal. Ask AI to help refine that goal (which is, in itself, a goal). But get the goal for AI clearly defined first.
4. Remember "if you don't know where you are going, any road will get you there... lost". So it is with AI. AI will give you an answer. It will most likely not be 100% correct, but it may be close enough, but how can you tell? Always double check the answer.
5. Running multiple sub-agents may get things done faster, but will cost a many more tokens because every time sub-agent is spun up you will path the whole ingest tax. For long projects this can really add up. For most jobs, stick with one model.
6. For repeated activities (or ones you think will be repeated) have the model create scripts that both it and you can run. This will allow you to do the activity without the model.
7. Getting clever with prompts is usually not worth the effort. Todays models (as of Oct '26) are really good at interpreting your intent from your questions. With that said, you still need to be clear on what the goal of the task is and any parameters that the model should align to. Giving a 'one shot' prompt is a gimmick.
8. When prompting do:
   a) Set a clear goal for the model. Define when the task is complete and the model is done.
   b) Provide any parameters you want it to follow or adhere to.
   c) Explicitly tell AI what you don't want it to do. Know that it may still ignore you and do it anyway. So make sure you have a backup way to prevent it.
   d) Tell the model how you expect it to save its progress. This will also help with the model being able to "remember" where it left off.
   e) Ask the model to repeat back to you or summarized its understanding of the ask before doing any work.
   f) Have AI create a plan before any work is done so you can review the plan. And do the review. This will keep both you and the model aligned.
   g) Explicitly tell the model to interview you or ask you questions for clarifications.
   h) Tell the model to record the assumptions that it makes in its history. This way you can review it.
9. Don't chase parameters, quantization, test suites, etc. Find a model that works for you and what you want to do. Build your own series of "tests" and what "good" looks like. Compare models to that.

Application Tips

1. Basically all application development habits and process you use today are what we want the model to use.
2. Make sure the model creates a Makefile or other standard build file for the application. This way you can make changes and/or swap out the models and still receive consistent results.
3. Always tell the model to initialize a new git project in the local working directory. Even if you don't intend to have the project be replicated anywhere else. This provide history of the work and the models generally provide really good summaries of what each commit provides.
4. For each feature or step in the plan, have the model create a new branch to do the work and when complete and tested, merge that branch back into main.

Final thoughts...

None. I will continue to update this list for myself as time goes on. Hopefully this will be helpful for you.