In partnership with

Prefer another language? Prefere receber em português? ¿Prefieres recibirlo en español? Choose here / Escolha aqui / Elige aquí: language form

Three cards for Microsoft Copilot, Google Gemini and Anthropic Claude, with what each is used for, and a fourth card saying one human checks every word, next to the title "The AI Trinity: when I use Copilot, Gemini, or Claude, and why".

I use three AI tools: Copilot, Gemini and Claude. But one human checks every word. Today I'll show you which tool I use for which job, and what happens behind these posts.

This post is a little different. I'll share parts of the process behind these posts, so you can see that there is a real person behind content made with AI.

I'll show you the challenges I had while building this process. The moments when I had to step in, because the AI was not giving me the result I wanted. The moments when I laughed out loud, alone, joking with the AI. And the moments when the checks I built made the AI itself warn me about a possible hallucination.

You'll see real screenshots from my conversations in the AI app. The exact words are there.

This time, I'm at my computer

As I have said many times here, the most beautiful part of all this, for me, is how much I'm learning. And I want to show other people that they can do it too, and learn a lot along the way.

Most of my posts start as voice notes, recorded on my phone, far from the computer. This one is different. I'm sitting in front of my screen, looking at the screenshots that will go into this post.

It's funny. I took these screenshots only this week, and this newsletter is less than a month old. But they still bring back memories. Human memories, not the AI's memory. Looking at some of them, I'm smiling alone again.

People who know me know I have a good sense of humor. Life in big companies is not easy. The higher you go in your career, the bigger the challenges. There's an old saying: the hard part is not getting to the top, it's staying there. Today, the hard part is staying in the big tech companies. I'm not complaining. It's natural. But everyone needs to find their balance, and mine is to laugh, many times at myself and at the things I do.

My drafts are not in order, and that's fine

Here is the draft of this post, exactly as I recorded it, in Portuguese. And below it, the same draft a few minutes later, pasted into the AI app with the title:

Top: the Portuguese draft of this post in Apple Notes, titled "The AI Trinity", dated November 6, 2026 at 21:01, one long block of spoken text. Bottom: the same draft pasted into the AI app with "New WTP draft" and the title, and the model picker showing Opus 5.5.

from my notes to the AI app

As you can see, it's one long block of text. I talk, I explain something, a few minutes later I remember something else and go back to it. I don't edit the draft, because that would only waste my time. The AI knows very well how to put everything in the right order. That's one of the ways it helps me. You can also see the model I use for this, at the bottom: Claude Opus 5.5.

And this is how the last post started. I sent the draft and a possible title, then the prompt I wanted in the post. The AI came back with the first draft, the small fixes it made, and three title options:

Top: the start of the chat for post 0011, "Let's start the next WTP draft, this Friday's tip", and the meeting prompt sent a few minutes later. Bottom: the AI's answer with the first draft, its small fixes and three title options.

the start of the last post

If you're curious about the whole process, I explained it in an earlier post.

The trinity: three tools, three jobs

I have tried many AI tools. Today I use only three.

Microsoft Copilot. Every day, all day, at work: my meetings, my architecture diagrams, and much more. The Clean Room architecture diagram I showed you in an earlier post was made with Copilot.

Google Gemini. Gemini and Google's family of AI tools are for research, analysis and videos. I'm not making many videos right now. I made more in the past, and I'll come back to it soon. I already have the whole sequence ready: the material goes in, and a finished video comes out. Gemini also works with the other Google tools I already use, and that matters a lot.

Anthropic Claude. Claude became the main tool for my personal projects, for one reason: my newsletter platform. I publish on beehiiv. It gives me the newsletter and the podcast you see here. beehiiv offers a connector that works with Claude, with step-by-step setup. That's why I started using Claude more often, and today almost my whole editorial and publishing process runs on it.

In short:

  • At work, every day: Copilot.

  • Research, analysis, video: Gemini.

  • The newsletters, from draft to publishing: Claude.

Someone may ask: "You work for Microsoft. Microsoft has all of this. Why use tools from other companies?"

Here is my answer. I started using them to learn what they offer. I also wanted to talk about Microsoft's products with a solid opinion. It's very easy to say good things about the product your company sells when you don't know what the competitor is doing. That opinion is not well grounded. So I started with Google's tools, then Anthropic's, to see what they are doing. I ended up liking them, and I kept them for personal use.

So I can say what each of the three does well. And honestly, all three have wonderful features for what you need them to do.

Yes, that Arnold Schwarzenegger.

He has a newsletter on beehiiv. So do Codie Sanchez, Jay Shetty, David Begnaud, Colin & Samir, and Joanna Stern.

They could publish just about anywhere. They chose beehiiv to build a direct connection with the people who want to hear from them most. And you can do the same, whether you’re world-famous or just at the start of your plan for world domination.

"If he's a beginner, why should we trust him?"

In the first posts, and in the podcast debate, I said I'm starting with AI tools. That's true. And the skeptical voice in the debate asks a fair question: if he's a beginner, why should we trust what he says?

When I say beginner, I don't mean someone who knows nothing. I mean a beginner compared to my level as a professional. As a user, I'm an advanced one. I know this tool very well. The proof is that I use it for almost everything in these publications, and it takes very little of my own time.

Now let me tell you how it's structured, why it worked, and what I learned.

Lesson 1: without rules, the files become a mess

The first big lesson: structure the process very well, and write rules that must be followed exactly.

I'm known for liking things organized. I pay attention to details, and I notice when something is done differently from one time to the next.

When I started with Claude, the results were very good. I keep all the files in a GitHub repository. Think of it as a digital filing cabinet that keeps every version of every file. I save the files on my computer first. Then I commit them, which is like taking a permanent snapshot of my work, and send them to GitHub, in the cloud. But after a few posts, I started to see a mess. Some files had one name, others another. Some were in one folder, others in another. The first post had one structure, the second had another, the third was different again.

When I saw this, I said: no, this won't work. I work like a production line. If every post is done the same way, my review is easy and fast. If every post is different, I lose time on every review.

So we wrote the rules. Today every original draft has the same name pattern, and every post folder has the same structure, from the images to the final text:

Two VS Code views of the repository. Left: the original-draft folder, one file per post, from post 0001 to 0011. Right: the folder of post 0011, with the same subfolders and files as every post.

the drafts folder and a post folder

When I see duplicated work, I stop it. For a while, every post doc existed twice, as a Google Doc and as a file in the repository. One question, one decision, and now there is only one:

The author asks whether the Google Doc and doc-01.md have the same content, then says to stop the Google Doc. The AI confirms the doc-0N.md files are now the only post doc.

one post doc, not two

The same with tools. One day, the AI started converting the podcast audio to transcribe it, in a different way than before. I stopped it and asked it to check what we already had:

Top: the author stops the AI, "Let's stop, I want to standardize this process", and asks it to check for an existing tool first. Bottom: the AI answers, "You were right: no WAV step is needed. Nothing like this existed in the repo yet."

let's stop and standardize

Nothing existed yet, so now there is one standard tool, in the same place as all the others:

The tools folder in the repository, with the new transcribe.py next to the other tools, such as utm.py and social_image.py.

the tools folder

The AI is fantastic. But if you don't write the rules for how you want things done, it may still give you the result, just not organized the way you need. And when you have to go back to that result later, you'll have a hard time.

Lesson 2: the cheaper model skipped steps

The second lesson was about quality, and about cost.

I pay for these tools, so I watch my credits. A common suggestion is to use a smaller, cheaper model for simple tasks and a bigger one for complex tasks. I won't even call it a recommendation. Think of a scooter engine and a truck engine: the scooter is great for quick trips, but for heavy work you need the truck. It makes sense. I thought: the whole structure of the post is ready, we built it together. From now on the AI only needs to follow it. That's simple, right? A smaller model should be enough.

In my experience, it wasn't.

The first time I tested the smaller model, Claude Haiku 5.5, I noticed right away. The quality was not the same. To give you a generic example: where I used to get five lines, I got two. We know a smaller model processes less, so that part I expected. What surprised me most was the execution.

My process has about 20 steps, from the draft to the published post. With the bigger model, Claude Opus 5.5, I say "run steps 2 to 7". It runs them all and stops at step 8, because step 8 is my review. With Claude Haiku 5.5, the same request stopped at step 3 or 4.

So I asked: why did you stop? Was it because I changed the model? This is the real message I wrote when I noticed it on one of my posts:

The author's message: the AI missed some steps, maybe because a different model was used at the beginning; the steps must be written in sequence, with done and pending steps shown every time. The AI rewrites the procedure as 20 numbered steps.

you missed some steps

That message is why the process is now 20 numbered steps, in order, and why every answer about a post ends with two lists: what is done, and what is pending, and who does it:

The end of an AI reply for a DBA post: the rules that apply, then a Done list with check marks and a Pending list with who does each step.

done and pending

And the steps are mine. When I want a different order, I change it, for this post and every post after it:

The author changes the order of the steps for this and all future posts, and the AI shows the new order, from step 10 to step 20.

changing the order of the steps

I tested the smaller model many times, and it happened again and again. This is my experience, and it may not be true for everyone. But it happened, and it still happens, in my case. If you have a tip on how to avoid it, reply to this email and tell me.

And if you think about it, you'll find the same behavior in people you work with, or even in your own family. Sometimes the AI behaves the same way: it simply stops, when it should have continued.

Lesson 3: when the AI called a true detail "invented"

Some people, when they see content made with AI, stay only on the skeptical side. They think it was all made artificially and may be full of hallucinations. That can happen. In most cases we can't avoid it 100%, and that's why review is always important.

But you can reduce it. One of the best ways is to write clear rules: what you want done, and what can and can't be generated.

Here is an example. The podcast debate about my posts is made by AI, with NotebookLM, a Google tool. Before I publish an episode, Claude checks what the hosts say against what I really published. This time, the check found a real error:

The AI's verdict after checking the podcast episode against the published posts: publish an October review, but not this version, because it gets the Bastion story wrong.

the verdict on the episode

The hosts got one of my stories wrong, so that version was not published. That is the check doing its job.

But in the same review, it flagged something else as "one invented detail". At the top, the AI's flag. At the bottom, my answer, and the AI's:

Top: the AI flags "one invented detail in post 9", the hosts saying "official PDF documentation". Bottom: the author explains he keeps a local folder with Microsoft docs in PDF from the official website, and the AI answers it was wrong to call it invented.

one invented detail, and my answer

From the AI's point of view, it was right. The hosts said I check technical points against "official PDF documentation", and my post only said "official documentation". But it is true. I keep a local folder with Microsoft's documentation saved as PDF, generated from the official website. I did it, but I never wrote it down. So the check saw something that didn't match what it knew, and flagged it as a possible hallucination.

When I read it, I thought aloud: I forgot to give it the right rules. Oops, I didn't tell you. From then on, we added it to the rules. Now it's official, and it won't be flagged again.

What I like here is simple. With these rules and protections, one tool can catch what another one did. A Google tool makes the debate, and Claude checks it against clear rules that must be followed. If one goes off the process, the other should catch it. If it doesn't, it means I configured the whole process badly. In the end, I'm responsible for the configuration.

The funny moment: "That's the reason I am paying you"

This one made me laugh a lot, alone, in front of the screen.

I decided to add a glossary at the end of every post, to make them easier for readers who are not technical. That was my idea from the beginning, but I had added it to only one post. When I noticed, I told the AI: a glossary in every post. The published ones, the current ones and the future ones.

The AI answered: "This is a big job." And I answered back:

The author asks for a glossary in all posts. The AI says it's a big job. The author answers: "yeah, a big job, that is the reason I am paying you, and by the way, you are doing great". The AI thanks him and continues.

"Yeah, a big job, that is the reason I am paying you, and by the way, you are doing great."

You know, like a boss: I know it's a big job, but you're doing great. Here's a little more, and we didn't see that, right?

Working while I sleep

The process is well structured. I start it on a computer dedicated to this, and it keeps running. Sometimes it takes a while to chew through all the data. There are the transcripts of the audio I generate. There are the checks I told you about: the check of the check of the check. But it all runs on its own.

One evening, I asked for the transcript of a 16-minute audio critique of my October posts, with the list of every recommendation. Then I went to bed. On the left, what it did. On the right, my phone the next morning:

Left: the AI's answer in the chat "October review critique transcript", the transcript of a 16:45 recording saved, followed by the recommendations. Right: the phone's Notification Center with the same message from the AI app, "Yesterday, 21:32".

Look at the time of the notification: 21:32. I took that screenshot the next morning. When I woke up, I picked up my phone to see my notifications, as I usually do. I always look for the ones from the AI app, to know whether it reported an error or finished with success. This time it had finished with success. I acknowledged it, and I knew the work was done, without me in front of the computer.

The scheduled tasks behind it all

Finally, what really makes all of this work when I'm not around: scheduled tasks. Here are some of them. I have many more than you see on the screen:

The Scheduled tasks page in the AI app: twelve active tasks, such as post reminders, ad results, a weekly subscriber language check and podcast review reminders. Some descriptions are blurred.

scheduled tasks

Some of them the AI created itself. It asks me if I want a reminder, or a check on something, and I say yes, go ahead, create it.

Here is a funny fact, or maybe not so funny. I get so many suggestions to schedule tasks from the AI that sometimes I wonder: are you scheduling tasks just to use up my credits? If that's it, dear AI, we'll have that conversation again, and I'll have to puxar a sua orelha. In Brazil, "pulling your ear" is a friendly scolding. Or maybe I'll pull your bits.

Try this week

  1. Write down one rule. Pick something you ask AI to do often, and write how you want it done: names, format, order. Give it the rule every time.

  2. Test the cheaper model on your real steps. If it skips steps or gives shorter answers, you'll see it fast. Compare before you switch for good.

  3. Add one check. Ask AI to compare its own result against your rules and flag anything that doesn't match. Then decide yourself.

Your turn

Reply to this email and tell me: what is one rule you will write for your AI this week? I read every reply.

See you next time.

Glossary

Term

What it means

AI model

The "brain" behind an AI tool. Bigger models usually handle more complex work but cost more to use.

beehiiv

The platform I use to write and send this newsletter, host its website and publish the podcast.

Claude

An AI assistant made by the company Anthropic. Opus and Haiku are two of its models: Opus is bigger, Haiku is smaller and cheaper.

Commit

Saving a set of file changes, with a note, into a repository's history.

Connector

A link that lets an AI assistant work directly inside another tool, like a newsletter platform.

Copilot

Microsoft's AI assistant. At work, it runs inside apps like Teams, Outlook and Word.

Credits

The usage you pay for in an AI tool. Bigger models and longer tasks use more of them.

Gemini

Google's AI assistant, connected to Google's other tools.

GitHub repository

A place where files are stored with the full history of every change. The local copy is on my computer; the remote one is on GitHub.

Hallucination

When an AI says something that sounds right but is not true or was made up.

LLM (large language model)

The kind of AI model behind tools like Copilot, Gemini and Claude. It learned from a huge amount of text, so it can understand and write language.

NotebookLM

A Google AI tool that works from your own documents. It can turn them into an audio conversation, like my podcast debate.

Notification

A short alert on your phone or computer, for example when a task finishes.

PDF

A file format that keeps a document looking the same on any device.

Scheduled task

A job that runs by itself at a set day and time, without anyone starting it.

Shell

A window where you type commands to the computer, instead of clicking buttons. One of the screenshots mentions it.

Transcript

The written text of an audio recording.

VS Code (Visual Studio Code)

A free code editor from Microsoft. Some screenshots show my repository in it.

Prefer to listen? Two AI hosts debate this post, one for it and one skeptic, on our podcast, We The People, in Tech or Not: The Debate: listen here

How I write: every post starts from my own notes and experience. I use AI tools to help draft and edit, and I review every word before it goes out.

Disclaimer: The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of Microsoft. The author is a Microsoft employee.