How To Get The Magic Thinking Rock To Do What You Want It To

The Backstory

A computer, fundamentally speaking, is a piece of Silicon. Number 14 on the periodic table, nested right between Aluminum and Phosphorus. Arguably two more important pieces of rocks, but I digress.

Every computer chip in existence is created from semiconductors. For the non-technical, pieces of Silicon we blasted with lasers to trick the rock into doing math. Add AI on top of that, and the math-doing rock can now help you write code, blog posts, translate things, come up with a workout program, or even help you get a raise. (Which you can help me to do as well, share this blog post so I get a lot of traffic to the website and can ask my boss for a raise, pls and thank you).

In one of my previous jobs as a Software Engineer, my boss had a really nice saying whenever something didn’t work: “The machine did exactly what you told it to do”. Which made sense at the time, software development is always deterministic. 2+2 is always equal to 4. Or 22, if you can’t do string to int conversion.

But then Sam Altman decided to release ChatGPT and now I have to explain to my parents what an “AI generated” video is, and that, no mom, the video of Donald Trump rapping isn’t real.

AI, by its very nature, isn’t deterministic. Not for most people at least, unless you already have a technical background or spend more time learning AI than any sane person should.

More to the point, AI changed how people work and even interface with computers. Openclaw, Hermes, Cowork, coding CLI tools, email filters, Jira AI assistants, I could go on for hours. Maybe I should, try to get our website ranking for some more keywords, but I do actually want people to read this, as I wrote this by hand. Imagine that, a gluten-free, gender-neutral, hand-written blog post in the great 2026, ikr?

One of the reasons why that is, is that I cannot use AI for content writing to save my life. Or, could not, but that’s why I’m writing this post, stick with me. The magic thinking rock can never seem to understand what I want it to do. Or it does understand, but decides that it won’t do it.

PICNIC errors aside (Google the joke, I won’t explain) the problem stands. I don’t face this issue with other things I use AI for, such as coding agents. Which of course triggered my engineering god complex and led me down several very, very deep rabbit holes, experimentation, and 3 in the morning rage-programming sessions.

I should be able to figure this out, right? I mean it’s not rocket science, it’s a blog post that I’m asking it to write. Talking to others in the industry, colleagues, friends, technical and non-technical people alike, I notice that I’m not alone in that.

Sure, AI does generate a blog post, but the blog post is at best meh. Sure, AI can edit a spreadsheet for you, but it doesn’t quite get what you want to do.

Why This Happens

For better or worse, AI is trained on the entire internet. Every piece of data it can gobble up makes its way into the training data for the annual Silicon Valley “Who has the better AI model” measuring contest. The majority of the players use synthetic data as well, but that’s beyond the scope.

Some companies even resort to pirating R-rated content off the internet to use in its training data. (Allegedly*)

But, the thing is, across any industry, across any specific skill, the vast, overwhelmingly vast majority of content is written for beginners. In many cases, also by beginners. This is natural and to be expected. The majority of software engineers at any given time are juniors to mediors. New grads are coming out, older software engineers are retiring. Same across the board — composers, writers, artists, etc. Even more so if the industry is in demand.

If you’re a content creator, or want to become one, and you see that “How to get a job in tech” is trending on Google, you follow the money, start creating content around that niche, monetize the traffic, profit. Business man doing business.

This leaves us in a pickle. Let’s say we disregard copyrighted works, you know, set an example for Fortune 500 companies — the majority of the entire internet traffic is social media and streaming sites. YouTube, Netflix, Instagram, etc. This means that AI is trained on a majority of beginner-level content. Which shows in benchmarks, specifically looking at HLE (Humanity’s Last Exam) where the best models sit right around 50%.

The irony of me writing this post for beginners is not lost on me.

There’s nothing wrong with being a beginner in any field, far from it.

But the fact is, if you’re trying to make heads or tails of an excel spreadsheet that your company has used since the early 2000s across every single Office version, good luck with getting AI to understand that. If you’re trying to get AI to create a high-ranking, high-converting piece of copy in your brand’s style and tone, good luck. Those things are hard to accomplish. If they were easy, good copywriters wouldn’t be worth their weight in gold. Good SEO people, good developers, you get the point.

What This Article Aims to Do

This isn’t going to be a prompting guide. I’m not going to give you a “READY TO USE PROMPT TO 100X YOUR AI PRODUCTIVITY”, as some clickbait YouTubers would like you to believe. The magic prompt, much like the fountain of youth, is a myth. And before you ask, yes, I am being paid to tell you that the fountain of youth is a myth.

In this article, my aim is to introduce you, (yes, you specifically reading this while on the toilet) to some techniques, technologies, mindset shifts that can further help you get more from your AI journey.

#1 – READY TO USE PROMPT TO 100X YOUR AI PRODUCTIVITY

See what I did there? Of course you do, you smart cookie you.

I’m not sure if you know this, but AI has spawned a bunch of new roles that haven’t previously existed. One of those, and I think maybe even the fastest growing one, is Prompt Engineer. Over 1000 open positions when I searched LinkedIn just now. A person who gets paid all day to come up with the best prompt possible to get the magic rock to do what he or she wants it to do.

Obviously more goes into the job, like different techniques, templating, tradeoffs, managing API costs, etc, but you get the point.

Good prompting is a skill. A high value skill, that can get you paid handsomely. Like, 6-figures handsomely. For me at least, you couldn’t pay me enough to sit around and argue with a sassy thinking rock for it to tell me what I want to hear for 8 hours a day, but hey, to each their own.

In case you’re like me, or don’t want to do a sudden career change, this leaves you with only one option. Learn to prompt. Google around (or bing around, I’m not judging), learn about the techniques, the differences, the methodology, everything. It’s a transferable skill between projects, clients, everything. It’s a good tool in your toolkit that will make a world of difference.

#2 – Rock picking 101

Much like engagement rings, the magical thinking rocks come in several different shapes and sizes and colors and cuts and quality and decorations and stories and so on. With companies racing across the globe to come up with the best thinking rock, a natural difference occurs. From a different methodology, values, training data, regulations, etc, the rocks come out with different strengths and weaknesses as a result. This is why, we as humans, while we still remain in control, came up with different benchmarks to evaluate what rock is good at what.

These are called AI benchmarks, a set of standardized tests that rank each model across a different set of criteria. From coding, to general intelligence, content writing, understanding, creative writing, etc.

The amount of benchmarks that exist are probably measured in the thousands. However, there are approximately <100 widely cited benchmarks. Those, while also serving as door frame to measure how much the little rocks have grown over the past year, have actual, real world value, depending on the task.

Learning how to read these, and understand them as well, will leave you with massively better performance. Instead of trying to jam a square peg in a round hole, you will instead be trying to jam a square peg in a square hole. Instead of trying to get one model to do something it may not be good at, it’s way easier and simpler to find a rock that just does it better.

#3 – Arming the rock with a Swiss Army knife

An AI harness, much like the ones worn by humans or animals, is a set of tools, of guardrails, skills or similar that wrap around a model and extend its capabilities. Like a tool belt around an electrician, so to speak.

You’re probably using some of them already, but they haven’t been named as such. Claude Cowork is a harness around Claude models. Codex is a coding harness around ChatGPT. GPT Researcher is a harness around any model.

Skills are a set of instructions that a model should follow. ELI5 — write what you want the model to do once, save it as a skill, and tell the model to use that skill to accomplish what you want. Or use skills someone else has written, that can also work.

These are important in several different ways. Using prebuilt skills made by other people who already had the pain of not being able to get the rock to understand what they want to do have gone through the trouble of writing a set of detailed, line by line instruction sets on how the rock should behave. There are many different skills repositories on the wider internet, but the one that makes most sense to start with is probably skills.sh. The most widely used one, and the one that has the most integrations, allowing you to easily install a custom skill exactly where you want it.

Alternatively, if you go through the exercise of writing painstakingly detailed, line by line instructions on what you need to get done, go through a couple of rounds of iterations, you will only need to do that exercise once, then direct the rock to use those instructions on every subsequent run.

Combining this with a harness is where the magic happens. Installing skills to Eigent, OpenWork, Claude Cowork, any coding CLI will literally make it have superpowers. I guarantee it.

#4 – Feeding the rock more data

Based on some statistics I saw on some YouTube short while doomscrolling, the average frontier AI model is trained on the same amount of data that a 5 year old toddler experiences from his surroundings. Obviously toddlers aren’t smarter than ChatGPT, and obviously ChatGPT is trained on the data, while we as puny humans only experience that amount of data from our senses.

You can imagine AI as a horse with blinders. No depth, just what’s in front of it, with no wider context. We, as humans, or the majority of us I should say, use our brain quite often without thinking. You probably don’t talk to your mom the same way you do with friends. Or, probably not to your boss the same way as a colleague.

When you’re writing an email to the person who signs your paychecks, you probably spend some more time on the email than to a colleague. Consciously or not, it probably happens to some extent. Or, if you have a cool boss like I do (I do really want that raise I mentioned at the start of the blog, help a brother out, share the post) you may write a joking email to him.

And let’s say that you ask AI to write an email for you. It doesn’t know the intricacies of the relationship. It doesn’t know who it’s for, what it’s about, what jokes to make and not to make. You’ve asked it to create an email, and it did. And you read the email and it’s empty, soulless, cold. Same for a piece of content. The rock doesn’t know what corporate tone you use, what jokes you like to make. The developer flavored rock doesn’t know that you prefer an arrow function to a regular one.

If you want the rock to replace you at work while you collect the paycheck, you need to feed it data, and a lot of it. Depending on the objective and goal, there exists a thousand and one tool. Here are some concepts to get you started.

RAG – Retrieval Augmented Generation

Rock-frienly database. You ask the rock a question, the rock searches the database for similar occurrences, semantic intent, analyzes that before responding to you.

Fine tuning – Reshape the rock into a more useful shape

The rock you have is circular. It does a bit of everything, but you need a tip for your spear. You reshape the rock to do the job of a spear tip, and do it exceptionally well.

Technical explanation, thank you google for coming up with AI overviews and saving me an extra click, but costing the website owner who originally wrote it traffic and potential monetization opportunities: Model fine-tuning is a machine learning technique that adapts a pre-trained model to perform a specific task or adopt a particular domain, style, or behavior. Instead of training from scratch, fine-tuning slightly adjusts the model’s internal weights using targeted, high-quality data to make it an expert in one focused area.

Email, Snail mail, Rock mail – Give the rock access to your emails

Inboxzero is one, I’m not sure about others since this one worked out of the box and I didn’t bother testing anything else.

#5 – Discord rock, telegram rock, whatsapp rock

Opening up a browser, navigating to ChatGPT, starting a new chat, explaining what you want it to do, uploading files, is such a bore. Since short form content collectively fried all of our attention spans, and I haven’t made a joke in about four paragraphs, as one of the most recognizable faces on the internet would say, “Ain’t Nobody Got Time for That”.

This is where agent gateways come in. Most often found in tools like Hermes. These effectively allow you to have an agent, or multiple, running 24/7 around the clock. These allow you to communicate with your agents via any and all communication platforms, at any and all hours of the day.

If you’re the type of busy CEO person who works at all hours of the day, this one is perfect for you!

If you’ve gone through the trouble of implementing my previous steps, recommendations and learnings, this is where the superpower starts kicking in. If you’ve created a dedicated SEO agent, content agent, development agent, and have all of them running behind some unified agent gateway, you’re truly, completely, location agnostic, model agnostic, the works.

Conclusion

After 7 whole pages of writing, I’m finally at the end of the tunnel.

The point I’m trying to make would effectively boil down to this — magic thinking rocks are cool, but they are not Mind Reading Magic Thinking rocks. Those are billed separately.

You can get the rock to do what you want to do, but it’s an investment. There isn’t an off the shelf solution. AI needs to be adjusted, tweaked, and thought through how to best serve the people in the environment it’s in.

Or, I mean, you can always pay us to do it.

LIMITED TIME OFFER AVAILABLE ONLY FOR THE NEXT CALLER.

Nah, I’m joking, but we do use everything described here in our day-to-day to bring maximum value to our clients and solution partners.

If you want to learn more, the CTA should be immediately below this paragraph, unless I forgot to close a p tag somewhere, in which case the entire page is probs broken, but I’ll blame the magic rock for breaking it

Thanks for coming to my TED talk.