Latest / We The People, in Tech or Not: The Debate / The AI Trinity: When I Use Copilot, Gemini, or Claude and Why

Transcript

The AI Trinity: When I Use Copilot, Gemini, or Claude and Why · We The People, in Tech or Not: The Debate

• 0:00
  Welcome to the debate. You know, specifically We The People in Tech or Not, the debate.
• 0:05
  And we should explicitly say right up front that this is an AI-generated podcast based on a post from the We The People in Tech or Not newsletter by Alex Rosa.
• 0:13
  Right, and its goal is to help listeners learn new ways to explain technical topics, going all the way from a beginner non-tech understanding up to an expert level. Exactly.
• 0:23
  So today, we are looking at Alex's post called The AI Trinity. And, uh, I want you to imagine hiring a brilliant, highly recommended executive assistant.
• 0:31
  On their first day, you hand them a stack of messy documents and say, "You know, please organize these." Sounds like a dream, honestly. Right.
• 0:38
  [laughs] But you come back an hour later expecting a neat desk, and instead, half the files are renamed in a language you barely understand. Some are duplicated for no reason, and others are just...
• 0:50
  well, they're shoved into entirely new randomly generated folders. Yeah, it completely shatters that mainstream illusion of, like, instant magic. Suddenly, you aren't managing a simple task.
• 0:59
  You are managing a very literal, very stubborn machine. Exactly. You realize this assistant isn't plug and play.
• 1:06
  If you want the job done right, you have to write a comprehensive ironclad manual detailing exactly how they should work, and I'll be defending that rigorous rule-based approach today.
• 1:15
  Applying strict engineered processes is really the only proven way to turn these AI tools into a reliable production line. And I will be taking the role of the respectful skeptic today.
• 1:25
  I'm approaching this from the perspective of the busy manager, the person who doesn't write code for a living. I mean, does getting reliable value out of AI actually require treating it like a rigid software project?
• 1:37
  I argue that it does. The sheer structure of this workflow is what makes the automation possible.
• 1:42
  But I feel like maintaining a 20-step checklist, managing multiple tools, and constantly debugging logic is just an exhausting, highly technical burden.
• 1:52
  It introduces unnecessary complexity that defeats the whole purpose of having an intuitive assistant.
• 1:57
  Well, let's lay out exactly what the author built because the success of his system kind of proves the necessity of the structure.
• 2:03
  He deliberately splits his tasks across three specialized tools, which he calls the AI Trinity. Right. Copilot, Gemini, and Claude. Yes, exactly.
• 2:11
  He uses Microsoft Copilot every day for corporate meetings, then Google Gemini for its deep integration with Google Tools for research and video, and finally, Anthropic's Claude to drive his entire long-form publishing pipeline on the Beehiiv platform.
• 2:24
  By enforcing strict rules across those three, he created a flawless digital assembly line. You call it a flawless assembly line, but, uh, let's look at who built it. The author is a tech-savvy Microsoft employee.
• 2:36
  He isn't just opening up a friendly chat window. He is deeply embedded in developer tools. He's using the right tools for the job, though.
• 2:44
  Sure, but he references using VS Code, which is a professional code editor, and he stores files in a GitHub repository.
• 2:50
  Uh, for anyone listening who doesn't know, a GitHub repository is basically a digital filing cabinet that keeps a permanent history of every file change. Right, which is incredibly useful.
• 3:00
  Useful, yes, but he's actively taking permanent snapshots so that when the AI inevitably breaks something, he can roll it back.
• 3:07
  He had to build rules just to stop the AI from duplicating files between Google Docs and his repository. Which is a brilliant proactive way to maintain version control. You know?
• 3:16
  He recognized AI can overwrite data unpredictably, so he built a safeguard. The AI takes his scattered Portuguese voice notes and perfectly formats them into English drafts. That part is definitely impressive.
• 3:27
  Yeah, it creates social media images, formats the text, but the AI can only do that heavy lifting because the human laid down exact unbending rules.
• 3:36
  The philosophy is one human checks every word, but the machine executes the labor. If you want real automation, you just have to build the factory floor first.
• 3:44
  But building a factory floor makes sense for a software developer. What about a marketing manager or an HR director?
• 3:51
  If you are a non-technical worker being forced to use developer environments and build 20-step checklists just to keep your files safe, I mean, that is an incredibly high barrier to entry.
• 4:01
  And it takes some learning, sure. It fundamentally flips the dynamic. The tool is no longer serving the user. The user is serving the tool, managing its quirks just to keep it on track.
• 4:11
  I think you're viewing the required structure as a burden rather than a mechanism for leverage. Let's look at the underlying compute before we dismiss it.
• 4:18
  This brings us to a major piece of his framework, model capability versus cost. Ah, right. The scooter versus the truck analogy. Exactly. It's a fantastic analogy.
• 4:27
  Smaller, cheaper models are like scooter engines, perfect for quick trips. But premium larger models are truck engines, and you absolutely need those for heavy work.
• 4:35
  It's a great way to visualize it, I'll admit, but it glosses over the actual user experience of driving that vehicle. Well, the data from his experience is incredibly telling about why this matters.
• 4:45
  He built that 20-step numbered checklist for his newsletter. He figured since his instructions were clear, a cheaper model should handle it, so he tested Claude Haiku 5.5, the scooter. And it crashed and burned, right?
• 4:57
  Literally stopped working at step three or four. It just gave up.
• 5:00
  But when he switched to the premium model, Claude Opus 5.5, he could command it to run steps two to seven, and it flawlessly executed them, stopping exactly at step eight for his manual review.
• 5:10
  Let's unpack the reality of what's happening financially here, though. You aren't just picking an engine. You are paying for credits, which is the usage currency of AI.
• 5:19
  Bigger models consume more credits, which costs real money. Right. Compute isn't free. Exactly.
• 5:24
  So if you hand a well-documented checklist to a system and that system just randomly clocks out after three steps without finishing the job, you have an unreliable service.
• 5:34
  We shouldn't praise a system where the budget option simply refuses to do the work it accepted. If I hire an assistant and have to watch them so they don't stop working halfway down a page, that isn't automation.
• 5:44
  That's micromanagement. But it's not about the AI refusing to work. It's about the physical limitations of the technology. These systems process information based on their context windows and working memory.
• 5:55
  A smaller model literally loses the thread because its computational capacity is maxed out. Which forces the user to babysit it. No, it forces the user to understand the tool.
• 6:05
  The author notes he expected shorter text from a smaller model, like two lines instead of five. What he learned was that execution logic also suffers when compute is low.
• 6:14
  That's not a flaw, it's just learning the mathematical specifications of your equipment. But look at the lengths he has to go to manage those limitations.
• 6:22
  Even after upgrading to the truck model, he literally had to write a permanent rule demanding the AI state the steps done and steps pending at the end of every single output. That's just to keep it on track.
• 6:33
  He forced it to generate a constant status report because he couldn't trust it to maintain its own place in the workflow.
• 6:39
  If I am paying premium rates, I shouldn't have to build a surveillance system to ensure it doesn't wander off the job.
• 6:45
  You're framing a tracking mechanism as surveillance, but in systems engineering, that is just standard operational logging. By forcing the AI to output its status, he actively anchors the model's attention mechanism.
• 6:56
  It's a technical workaround that forces the AI to reread its own progress. And my counter is that the average professional shouldn't have to hack an AI's attention mechanism just to get a newsletter published.
• 7:07
  This constant need to engineer around flaws leads directly to the most exhausting part of this framework: the rules, the hallucinations, and documenting reality.
• 7:17
  I completely disagree with the premise that this system is fragile, actually. The system's ability to catch errors proves its robustness. He implemented what is essentially a check of the check. Using NotebookLM, right?
• 7:28
  Yes, a Google tool called NotebookLM. He generates an audio debate about his written posts. Because he knows AI can hallucinate facts, he doesn't publish blindly.
• 7:37
  He takes the Google audio transcript, feeds it into Claude, and commands Claude to cross-reference it against his original text. And we should definitely note that it did catch a real hallucination there. It really did.
• 7:48
  The audio host confidently told a completely wrong story about a Bastion code. Claude successfully flagged it. This is a self-auditing digital factory.
• 7:56
  When you have different models checking each other based on strict rules, you eliminate the single points of failure. The Bastion code catch was great, but the rest of that interaction reveals a deeply flawed paradigm.
• 8:07
  In that exact same audit, Claude aggressively flagged the phrase "official PDF documentation" as an invented detail. Yes, because of a minor technicality.
• 8:16
  A technicality that accused the author's own work of being a hallucination. The AI pointed out, you know, "Your original post only says official documentation. You never mentioned a PDF."
• 8:27
  Which was technically accurate based on the text the AI was fed. Technically accurate, but functionally absurd. The AI is more literal than a lawyer now.
• 8:35
  The author realized he did generate those PDFs locally from the web, but because he never wrote his personal local folder habits down in his master rule book, the AI flagged reality as a hallucination.
• 8:46
  It just caught a discrepancy. But it forces the human to constantly debug the AI's logic. If your undocumented reality doesn't match the AI's master file, the AI assumes reality is wrong.
• 8:57
  He literally had to stop his workflow and go update his rule book just so the AI would stop throwing red flags at accurate information.
• 9:05
  The fact that it flagged a discrepancy means the guardrails are functioning exactly as designed, though. The AI doesn't have eyes, it only has the rules.
• 9:12
  As the author explicitly notes, if a discrepancy slips through, it means he configured the process badly. He says, "In the end, I'm responsible for the configuration."
• 9:20
  That is an immense amount of responsibility for a casual user. But that level of absolute control and accountability is what professionals need to ensure quality output.
• 9:29
  You are looking at a highly effective quality assurance system and dismissing it as a bureaucracy. I am dismissing it because it requires a human to document every single minor action they take.
• 9:39
  The author even laughs about it, telling the AI, "That is the reason I am paying you." It's a lighthearted moment, but it highlights immense upfront friction.
• 9:48
  Does a busy marketing director genuinely have time to update a master rule book every time they change a minor habit? You endure that upfront friction for the ultimate payoff, which is true asynchronous automation.
• 9:59
  Let's look at the screenshot the author shared from twenty-one thirty-two in the evening. The overnight processing? Exactly. He had a sixteen-minute audio critique of his October posts.
• 10:08
  He didn't sit down and type it out. He asked the AI to extract the transcript, format it in Markdown, and list all actionable recommendations. Then he simply went to sleep.
• 10:17
  I will admit, the idea of the processing happening overnight is undeniably impressive. It's a complete paradigm shift.
• 10:24
  While he slept, the AI ground through the data, generated the file, and even used its contextual awareness to fix obvious speech-to-text errors.
• 10:32
  It intelligently changed the nonsensical phrase "pulse six" to "post six" and corrected P-E-G-O to the acronym P-A-Y-G. Right. The human just wakes up, checks a phone notification, and the work is perfectly formatted.
• 10:46
  The strict rules, the premium models, they all serve this single objective: separating the output from the human's immediate time and effort. The work happens without you in front of the screen.
• 10:57
  The narrative of working while you sleep is seductive, I mean, truly. But let's look closely at the actual human behavior described here. Is he actually freed from the work, or is he just carrying it with him everywhere?
• 11:08
  What do you mean? He explicitly notes that the very first thing he does in the morning is check his phone notifications to see if the AI reported an error or finished with success. That is an always-on mental load.
• 11:20
  You aren't separating yourself from the work. You are just moving the anxiety of task management from your office desk into your bed. Checking a notification takes two seconds.
• 11:30
  That is a massive upgrade from spending an hour manually transcribing audio. But look at how it's operating in the background. He isn't just clicking a button on an app.
• 11:38
  He has set up numerous scheduled tasks, and he's constantly managing shell commands. Shell commands are just text inputs. Yes, but it means typing raw backend text commands into a terminal instead of clicking buttons.
• 11:51
  He's essentially hacking his own operating system to keep the AI on schedule. The mental load of this is immense.
• 11:57
  He even jokes about the AI suggesting so many automated tasks that he worries it's intentionally trying to use up his paid credits. Right. He uses that Brazilian phrase.
• 12:06
  Yes, puxar a sua orelha, which means pulling its ear, giving it a friendly scolding. It's clearly a lighthearted joke. He enjoys building the system The joke masks a very real anxiety about runaway costs, though.
• 12:20
  He is literally worried his digital employee is creating busy work that will cost him real money while he's asleep.
• 12:25
  Do you really wanna start your morning checking terminal logs and worrying if your assistant is burning through your budget? That isn't working while you sleep. It's sleeping with one eye open.
• 12:35
  But the alternative is doing all the labor manually, which simply does not scale. The upfront investment in writing clear rules and paying for those premium Chuck models yields an incredibly powerful result.
• 12:46
  Yes, he had to learn how a GitHub repository works. Yes, he has to enforce a strict checklist, but look at the output. The output is good, yes. It's a professional multi-platform publishing empire.
• 12:59
  It transcribes audio, drafts newsletters, formats code blocks, and actively catches its own hallucinations, all run by a single human being who only has to review the final words.
• 13:09
  The friction you're highlighting is temporary, but the leverage he achieved is permanent.
• 13:14
  I just maintain that the barrier to entry for that leverage is effectively taking on a second job as a systems administrator, meticulously managing the trinity of Copilot, Gemini, and Claude, and building an exhaustive checklist just so the AI doesn't quit halfway through.
• 13:28
  You've transformed an intuitive tool into a demanding software project. For the non-technical manager, the cost is simply too steep.
• 13:35
  Well, despite our deep disagreement on the mental load, there is a profound point of convergence here. It lies in the author's very first foundational lesson.
• 13:44
  I think we both agree completely on this statement: without rules, the files become a mess. Absolutely. That is the undeniable truth of the current technology.
• 13:53
  Whether you build a 20-step digital factory or just want a simple draft of an email, leaving an AI to its own devices with zero structure leads to absolute chaos.
• 14:02
  We saw exactly what happened before he implemented his master rules. Inconsistent names, duplicated docs, unpredictable outputs. Exactly. Structure is not optional if you want quality.
• 14:12
  The debate we are having is merely about how much of that structure is practically sustainable for the average person. Which leads us to a highly practical next step for anyone listening.
• 14:20
  You do not have to go out and build a GitHub repository today. You don't need to write shell scripts. Instead, just follow the author's core philosophy on a micro scale.
• 14:29
  Pick just one repetitive task you do with AI this week. Just one. Just one. Take five minutes to write down exactly how you want it done, the specific formatting, the exact order of information, the naming conventions.
• 14:40
  Save that text, then feed that exact rule book to the AI every single time you ask it to do that task, and see how the consistency fundamentally changes your results.
• 14:49
  And if you are curious to see the actual mechanics of this entire process, the raw screenshots of his rules, the full 20-step checklist, the exact moment it flagged the Bastion code hallucination, and the prompts he used, you can find all of that raw evidence in the source material.
• 15:03
  We highly invite you to examine the evidence, see the screenshots, and read the full post for free at InTechOrNot.com. That is InTechOrNot.com.
• 15:12
  It is absolutely worth looking at, if only to understand the sheer amount of engineering that currently goes on behind the scenes of these seemingly magical tools.
• 15:20
  Because at the end of the day, when you hand that stack of messy documents to your digital assistant, you have a choice to make.
• 15:26
  Are you gonna cross your fingers and hope for magic, or are you gonna sit down and write the manual?