# How I Use AI on This Blog

AI interviews me, drafts, checks every sentence, and translates. Here is what each job does, who checks it, and why the checks matter more than the tool.

2026-10-12

AI will probably touch every post on this blog, as it touched this first one. That fact alone, I think, tells you little about the structure or the content of a post. What the models did, how they did it, and who checked it tell you much more about whether to trust it.

Amid this crazed hype train, being skeptical about AI is a fair position: if AI touched a text, why trust it? Large Language Models (LLMs), the specific technologies that support applications like ChatGPT, Claude and Gemini, can spew out complete nonsense in the most terrifying way possible: fluently and confidently. Even the big labs acknowledge it. \[1] \[2] \[3]

My point of contention is that the models themselves now matter much less than they did a year ago. What decides whether you get something useful or just more slop is the harness: the thing that sits on top of the models and handles everything else.

This blog is where I try to prove it.

It's a personal venture. I missed writing about stuff that interests me, particularly what I build and what I learn while building it. I've been building for the web since 2007, when I first discovered the power of CMSs like Joomla and Drupal, but only at the end of 2024 did I go all in on code. Right now, my two main projects are Desio, a platform that helps restaurants turn their clients into regulars, and kflow, the tooling that we use as a team to handle our day-to-day operations at TCOM, our parent company. The opinions here are my own, unless I say otherwise.

## The model matters less than what surrounds it

Most programmers use AI now. In JetBrains' 2026 survey, 90% of professional developers used AI coding agents at work every week, and 68% used them every day. \[4] Stack Overflow's 2026 survey shows the same habit: 73% of developers who use coding assistants use them daily. \[5]

Stack Overflow has asked a similar question every year since 2023, and the share keeps growing.

![The share of Stack Overflow survey respondents who use AI tools, from 44% in 2023 to 83% in 2026.](https://ricardokrug.com.br/en/blog/welcome-world/figures/ai-use.png)

My guess is that the real numbers are higher by now. Some companies no longer leave AI optional: Shopify made it "a baseline expectation" in 2025, \[10] and Meta said it would weigh it in performance reviews from 2026. \[11]

If you're not a programmer, you may not realize how differently programmers use LLMs, compared with a chatbot window (although, to be fair, that change is fairly recent too). Code may still be written in the same chat app you use to make funny images of your friends, but in my experience, the AI inside a business system is a completely different beast. That is what we call a harness: the code around a model. It decides what the model sees and which tools it can use, can check what comes back, and sends the result where it needs to go.

![An invoice harness: a workflow picks up the email, the model reads the PDF, a script checks the result, and the invoice goes to the accounting system or to a person.](https://ricardokrug.com.br/en/blog/welcome-world/figures/invoice-harness.png)

Picture a supplier emailing an invoice as a PDF. Nobody opens a chat window. A workflow picks up the email, hands the PDF to a model to pull out the tax ID, the items and the totals, and then a plain script checks the math and validates the tax ID. If everything adds up, the invoice goes straight into the accounting system; if anything is off, it lands on a person's desk. The model in that pipeline may be the very same one behind the chatbot, but the harness around it is what turns a good guess into something a business can rely on.

So I don't think "I use AI" tells you much about the work. It can mean a free chatbot account, which gets the vendor's smaller, cheaper models. \[12] \[13] It can also mean a setup like mine: models from different vendors, each in its own role, and scripts beside them that check what a model can't promise, giving me (ideally) the best of both worlds: models that make good guesses, and scripts that never guess.

Careless use of AI gives bad results. Careful use exists, but the label "I use AI" doesn't tell you which one you're reading.

Let's get back to the blog and the harness I built to write for it. For each job below, I show two things: what AI did, and who checked it.

![AI and I take turns on each post. Each AI step ends with my decision.](https://ricardokrug.com.br/en/blog/welcome-world/figures/ai-jobs.png)

## AI asks the questions, and I give the answers

Before I write a word, an AI interviews me. The idea comes from Matt Pocock's grill-me skill, which tells an AI to "interview me relentlessly" about a plan. \[14]

I adapted it for writing. My version asks eight questions in a fixed order, one at a time. Who is the target reader of this piece? What do I imagine they believe about the subject? And what do I claim that agrees with or counters their view? Each answer must pass a test before the next question, and the tests come from technical literature on argument and plain writing, such as Minto's *The Pyramid Principle*.

For each question, the model recommends one answer, usually with its reasoning for that "choice". In my experience, its recommendations are sometimes biased in ways I disagree with: models tend to favor their own output \[19] and to tell users what they seem to want. \[20] So the interview usually takes several revisions.

This post is a good example. I changed the topic twice during the interview, and the reader twice. The belief this post answers is one the AI didn't recommend. And I put back a point I had dropped: the translation, which you'll read about below.

That may seem counter-productive at first (or even a skill issue), but it isn't far from any regular ideation process without AI. Our first ideas usually suck. Hard. We may have a general idea of a topic, even glimpse a connection between it and another one. But as soon as we put pen to paper (or finger to keyboard, in this case), in my experience, even the slightest critical look makes us throw most of it out of the window.

I think iteration is the heart of any creative process. Or, to use probably one of the biggest clichés in quotable history:

> Genius is one percent inspiration, ninety-nine percent perspiration.
>
> — Thomas Edison

> \[!ASIDE]
> A fitting quote for a post about the powers of AI, by the way, because Edison was really a master of two things: inventing, and putting his name on other people's work. \[15] \[16] \[17] \[18]

Also worth noting: during this first phase, the model fills in no answer for me (the question/answer bias applies here too). I usually answer only a few multiple choice questions before I feel the need to express myself more verbosely. Sometimes I type, but most times I dictate my answers to yet another model to transcribe.

> \[!ASIDE handwritten]
> Yes, this is how I think, write and talk — I used to be a lawyer, what can I do.

## AI writes a first draft, and I decide what stays

So who actually writes the sentences? Often, at least at first, a model does. But it doesn't start from a blank prompt: it works from the brief that came out of the interview, and from a ledger of sourced claims — the facts I plan to use, each next to its source. That's how this post started.

I'll be honest: that first draft often ~~is absolute garbage~~ isn't good. It holds some of my ideas, sure, but it misses the flow of my thought — the order in which I'd naturally connect one point to the next, the detours I'd take, the jokes I'd make (bad ones, mostly). It reads like someone summarizing me, not like me.

So why bother? Because it gets me past the empty page. And for me, starting is always the hardest part. It has also been nice to notice how, every now and then, it brings up an idea I hadn't seen, or a connection between two points that makes the argument stronger.

For me, that's the real job of the draft. Not to be the text, but to show me a first preview of how the ideas link, what's missing, and what I assumed wrongly. It works a lot like those first ideas from the interview: most of it goes out the window, but I only find out which parts once they're on the page. Then I rewrite it, and I decide on each change.

![The "So why bother?" paragraph: the short AI first draft beside the version I published, with my additions highlighted and my cuts struck.](https://ricardokrug.com.br/en/blog/welcome-world/figures/before-after.png)

On this post, an AI pass proposed 79 changes to make the draft sound more like me. I kept 68. Deciding which ones to turn down is a call no model can make on my behalf.

## Two checkers read every sentence

Fine, the argument is mine. But are the facts right? That's the part where I trust myself the least. Not because I'm necessarily careless, mind you, but because it's hard to read your own text with fresh eyes.

So before a post ships, two models check it, one sentence at a time. One is GPT-6 Astra, from OpenAI. The other is Claude Opus 5.5, from Anthropic. Each one opens the source that a sentence cites and checks the sentence against it: does the page actually say that, with that number, in that context?

The rule is simple. A sentence passes only when both checkers pass it. If either one flags it, it comes to me. Then I either fix it, or I keep it and write down why. Again, this may seem like another waste of time, but I still insist on it: if you disagree with something, you should be able to show your reasoning. Models (may) find it later and (try to) respect it. I try to do the same with humans, although we seem much less moved by such silly efforts.

On this post, the checkers read all 168 sentences of the draft. And that wasn't the only round: every time I changed a set of sentences, they ran again on those.

![Three sentences from this post's checks: one both checkers passed, one that one checker flagged, and one that both flagged.](https://ricardokrug.com.br/en/blog/welcome-world/figures/check-outcomes.png)

## Two AIs can share a blind spot, so they don't decide

But Ricardo, you may ask, isn't that just AI grading AI?

I think that is a fair question, and one I asked myself more than once while building this. Two models that agree aren't proof of truth. Much of their training data likely overlaps (much of the same internet, many of the same books), so they can miss the same thing for the same reason.

That's why a flag never goes to a vote, and the checkers never settle it among themselves. A few things work in my favor here. They come from two different companies. They also hold each sentence against its source, not against their own memory, which is where models tend to make things up. And every flag comes to me, with the reason for each call written down.

Checking sources is old work for me. I come from a legal academic background, so I've done my share of reading. Quite a bit, as you might expect. I still keep many secondary sources around, both for my own fact-checking and, once digitized, for the models to consult.

![A study lined with old law books, with an open volume and a scanner copying its pages on the reading table. Illustration made with ChatGPT (gpt-image-2.5-flare).](https://tn0oilsmy7aw0f5x.public.blob.vercel-storage.com/posts/welcome-world/library.png)

The main point here is that the models do not decide what is true. Every fact needs to be properly sourced (and verified against that source). What they effectively decide is what I need to look at. That's a much smaller job, and one that many of them seem to be quite good at.

## AI translates, and a script holds the facts in place

Although I'm Brazilian, I've got to admit to something: I feel much more comfortable reading and writing in English. While I realize this may leave me in a precarious spot when I need to say certain things in Portuguese, most of the tech world's knowledge is still published in English first, in my experience. So, while I could write in Portuguese first, I honestly prefer to write in English.

I want every post in both languages. So is the other version just machine output?

Partly, yes. And I'd rather say it than pretend otherwise.

AI usually translates each post from the checked English (I may write some topics in Portuguese first in the future, and the harness can handle either order). I ask it to translate faithfully, in natural Portuguese, and to keep every claim, example and number. And even though I've talked a lot about harnesses so far, here is one place where the choice of model really matters. I started with GPT-6 Astra doing the translation (it is an awesome Portuguese typist, in my opinion). When Google released Gemini 4 Argon to the public on October 13, I tried it on this post (things really do move this fast, it is terrifying). In the end, I went back to Astra: I felt that the idioms were still more faithful in OpenAI's top-tier model.

![The survey paragraph in English and in GPT-6 Astra's Portuguese, with each number the script compares highlighted in both.](https://ricardokrug.com.br/en/blog/welcome-world/figures/translation-held.png)

After I finish my first review, a script holds the facts in place. It compares the two versions and refuses any translation whose headings, figures, links, numbers or tag count differ, and a test runs the same check on every published pair. This is the deterministic half of the harness doing its job: a script doesn't get creative, doesn't get tired, and doesn't round a number because it sounded better. So a pair that passes doesn't get fact-checked a second time: the facts were checked in English, and the script makes sure they made it across.

The wording gets a softer check. A second tool flags literal translations: an English phrase dressed up in Portuguese words (anyone who has read a badly translated manual knows exactly what I mean). On this post, a second model reviewed the Portuguese and found 7 issues. I fixed all of them.

The script has a limit, and you should know it. It checks structure, not meaning. A wrong word with the right numbers sails right through it. So the meaning rests on that second tool and on my own read — which is to say, on the same kind of review as everything else here (stop looking at that em-dash, it is mine, I've written it purposefully and I love it).

## The checks are slow, and that helps

Right now, you might be thinking: is this all even worth it?

I'll be honest about the cost. Every changed sentence goes back to both checkers, even after the smallest edit, like a comma or a swapped word. That can be extremely annoying, and some days it halts the work altogether.

And it's not only the waiting. Every round costs tokens (and therefore money) and, more importantly, attention: each flag has to be read and decided on, and each one I keep has to be justified in writing. Multiply that by every sentence I touch after the first pass, and a post that started on Monday and felt finished on Thursday can easily still be in review by the beginning of the next week.

In the end, I don't think that's a bad thing. I am an anxious person, courtesy of a late ADHD diagnosis. These rounds force me to pause, step away from the task, and come back to it later, whether I want to or not. And when I do come back, I read the text a little more like a stranger would. Then I re-read, reconsider, and push out something that actually makes sense.

![An empty desk with an open laptop, and through an open door, a person walking away down a sunny city street. Illustration made with ChatGPT (gpt-image-2.5-flare).](https://tn0oilsmy7aw0f5x.public.blob.vercel-storage.com/posts/welcome-world/step-away.png)

It's the same lesson as the interview and the draft, really: the first version is rarely the good one. In the interview, the model pushes back on my answers; in the draft, most of what it writes goes out the window; here, the checkers send me back to sentences I thought were done. The checks just make that iteration hard to skip (which, if I'm honest with myself, is exactly the step I'd be most tempted to skip on a busy day).

There's a lot of talk about AI slop: text that is fluent, confident and empty, produced faster than anyone can read it. I don't think the cure is to stop using the tools. To this day, the best antidote to AI slop I know is review. Then more review. Ideally by someone (or something) other than whoever wrote it.

## Judge AI by its job

So, after all of that, how should you read the posts here?

Not by asking whether AI touched them — it did, and it will. That question made more sense a few years ago, when using AI was a choice that said something about the author. Today, at least among programmers, it's closer to the default (the numbers at the start of this post show it), and "made with AI" tells you about as much as "written on a computer".

If you decide to indulge me and read anything else, judge every text here by its format, its content and its sources. When it comes to AI, ask what it did and who checked it. Did a model come up with the argument, or did it help frame it? Were the facts held against sources, or against a model's memory? Did a person make each call, and can you see why? That goes for this blog, and for anything else you read: a news article, a report at work, a post on social media. The process behind a text tells you almost everything; the tool, almost nothing.

For transparency's sake, every other post I publish here will link back to this page, or to its latest version. Now you know what each link stands for.

## References

1. Primary: Kalai et al., ["Why language models hallucinate"](https://openai.com/index/why-language-models-hallucinate/), OpenAI, September 5, 2025.
2. Primary: Anthropic, ["Claude is providing incorrect or misleading responses. What's going on?"](https://support.claude.com/en/articles/8525154-claude-is-providing-incorrect-or-misleading-responses-what-s-going-on), Claude Help Center, March 16, 2026.
3. Primary: Google, ["Learn about responses from Gemini Apps"](https://support.google.com/gemini/answer/16279220?hl=en), Gemini Apps Help.
4. Primary: Bogdanov, ["AI Coding Agents: Adoption Trends"](https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/), The JetBrains Blog, August 18, 2026.
5. Primary: Donovan, ["The results of the 2026 Developer Survey are here!"](https://stackoverflow.blog/2026/10/06/the-results-of-the-2026-developer-survey-are-here/), Stack Overflow Blog, October 6, 2026.
6. Primary: Stack Overflow, ["2023 Developer Survey: AI"](https://survey.stackoverflow.co/2023/#ai), 2023.
7. Primary: Stack Overflow, ["2024 Developer Survey: AI"](https://survey.stackoverflow.co/2024/ai), 2024.
8. Primary: Stack Overflow, ["2025 Developer Survey: AI"](https://survey.stackoverflow.co/2025/ai), 2025.
9. Primary: Stack Overflow, ["2026 Developer Survey: AI"](https://survey.stackoverflow.co/2026/ai), 2026.
10. Primary: Lütke, ["Reflexive AI usage is now a baseline expectation at Shopify"](https://x.com/tobi/status/1909251946235437514), X, April 7, 2025.
11. Press: Business Insider, ["Meta is about to start grading workers on their AI skills"](https://www.businessinsider.com/meta-ai-employee-performance-review-overhaul-2025-11), November 2025.
12. Primary: OpenAI, ["GPT-6 and other models in ChatGPT"](https://help.openai.com/en/articles/20001354-gpt-6-and-other-models-in-chatgpt), OpenAI Help Center, October 2026.
13. Primary: Anthropic, ["Plans and pricing"](https://claude.com/pricing), Claude, October 2026.
14. Primary: Pocock, ["grill-me"](https://github.com/mattpocock/skills/blob/78649baa3e6e4c241ce85ee8dc93ecc0b53e4515/grill-me/SKILL.md), mattpocock/skills on GitHub, February 25, 2026.
15. Analysis: Thomas A. Edison Papers, ["Biography"](https://edison.rutgers.edu/life-of-edison/biography), Rutgers University.
16. Analysis: Library of Congress, ["Origins of Motion Pictures"](https://www.loc.gov/collections/edison-company-motion-pictures-and-sound-recordings/articles-and-essays/history-of-edison-motion-pictures/origins-of-motion-pictures/), Inventing Entertainment collection.
17. Analysis: Ashworth, ["William Dickson"](https://www.lindahall.org/about/news/scientist-of-the-day/william-dickson/), Linda Hall Library, August 3, 2020.
18. Analysis: ETHW, ["Thomas Edison at Menlo Park"](https://ethw.org/Thomas_Edison_at_Menlo_Park), Engineering and Technology History Wiki.
19. Primary: Panickssery, Bowman and Feng, ["LLM Evaluators Recognize and Favor Their Own Generations"](https://arxiv.org/abs/2404.13076), arXiv, April 15, 2024 (NeurIPS 2024).
20. Primary: Sharma et al., ["Towards Understanding Sycophancy in Language Models"](https://arxiv.org/abs/2310.13548), Anthropic, arXiv, October 20, 2023.