Transcript 0:00 Welcome to We the People in Tech or Not: The Debate. 0:04 This is an AI-generated podcast based on a post from the We the People in Tech or Not newsletter by Alex Rosa, and its goal is to help listeners learn new ways to explain technical topics from a beginner, non-tech understanding up to an expert level. 0:18 Right. And today, um, we're really digging into this fascinating intersection of human experience and artificial intelligence. Exactly. We're looking at a very specific workflow. 0:29 It's designed to, you know, successfully hand off heavy, repetitive tasks to AI while keeping the human completely in charge of the ideas and the final decisions. Which is a huge promise, right? Mm-hmm. 0:39 I mean, the foundation here is this eight-step process the author uses. He takes these roughly sixteen hundred word Portuguese voice notes and turns them into highly polished English publications. Yeah. 0:50 And I look at this, and I just see, well, a perfect blueprint for any professional today. 0:55 It proves that AI can dramatically enhance our output without sacrificing those thirty years of unique lived experience that a professional brings to the table. See, I come at it from a slightly different angle. 1:05 The, the theory sounds incredibly appealing, especially for, like, a busy manager or a non-technical leader who, who's just trying to survive their weekly workload. Sure, yeah. 1:13 But I question the hidden complexities of this system. When you look closely at this eight-step process, you really have to ask, at what point does managing the AI take more effort than just doing the work yourself? 1:24 Oh, interesting. Yeah, and there's a profound risk of losing one's true, authentic voice in that translation. 1:31 Furthermore, I think we have to honestly evaluate whether the massive technical setup required here is truly practical for the average busy professional. 1:40 I completely see why you think that, but let me give you a different perspective on how we should frame this workflow. It fundamentally separates the generation of ideas from the, uh, the mechanics of publishing. Okay. 1:53 How so? Well, the process starts with the author speaking freely out loud in Portuguese, and these raw recordings, they range anywhere from eight hundred to over three thousand words. That's a lot of raw audio. 2:04 It is, and there's no structure yet. It's just the raw story, the memories, the lessons from real customer interactions or questions from colleagues. 2:12 And to really understand how this functions, I look to the ghostwriter analogy. The ghostwriter analogy? Right. Famous book authors, business leaders, they use human ghostwriters all the time. 2:23 The recognized author provides the ideas, the messy notes, the life experience. Yeah. And then the ghostwriter takes that raw material and structures it into polished chapters. 2:32 But the original person retains full authorship because the substance, you know, the soul of the work, is entirely theirs. Right. The core narrative. Exactly. In this workflow, AI is just acting as a digital ghostwriter. 2:45 The human provides the thirty years of hard-earned career experience, and the AI provides the formatting and the structural heavy lifting. That ghostwriter analogy, I mean, it paints a very clean picture, 2:55 but we really have to look at the real-world management required to keep a digital ghostwriter in check. What do you mean? 3:00 Well, for a busy non-technical manager, the magic of this workflow actually hides a tremendous amount of meticulous oversight. The author isn't just saying, "Hey, make this sound good." No, definitely not. Right. 3:13 The parameters demand, and I'm quoting here, "Correct, clear English, short sentences, a Brazilian voice, and no marketing tone." Keeping an AI constrained to those exact specifications does not just happen by default. 3:26 True. It takes setup. Exactly. How do you actually force an algorithm to maintain a Brazilian voice without it drifting into, you know, robotic corporate speak? It requires strict, continuous management from the human. 3:39 That is a crucial distinction, though. You don't achieve that by just typing a casual prompt into a chat window every time you want to write something. Right, which is what most people do. Exactly. 3:48 But the mechanism here relies on setting up a persistent set of instructions, uh, system prompts, that act as permanent guardrails. Okay. 3:56 You're programming the AI's default behavior so that before it generates a single word, it is forced to cross-reference its output against those strict rules. 4:04 No marketing fluff, short sentences, retain the cultural tone. 4:08 But setting that up- You do the heavy lifting of setting up those constraints once, just once, and then the system applies them to every future voice note automatically. 4:16 But even with those guardrails, we know the technology struggles with authenticity. I mean, the source material itself provides evidence of the struggle. You mean the feedback from his colleagues? Yes. 4:27 The author openly admits that recently colleagues gave him specific feedback on the language itself and how he uses AI. If English isn't your first language, even if you are entirely fluent, certain things stay with you. 4:40 Right. You build sentences differently. Exactly. You might not use certain English idioms simply because you don't know they exist, and native speakers notice this. 4:48 They might not be able to pinpoint exactly why, but they can tell when a sentence feels slightly off. Or too perfect, yeah. Right, when it feels too artificially polished to be natural. And to me, this proves my point. 4:59 A tool cannot perfectly capture human authenticity without leaving digital fingerprints all over the prose. 5:05 Well, I completely concede that AI isn't perfect at capturing voice out of the box and that it leaves those digital fingerprints. 5:12 But that is exactly why the author's workflow keeps the critical steps firmly in human hands. Which steps specifically? Step one is the unedited human voice note. The AI does not invent the stories, right? 5:24 It translates and drafts based only on what was spoken. And then we jump to steps five and six, the human reviews and the human decides. 5:31 But if the draft is full of weird idioms- That's when the author asks himself, "Is this what I mean? Is the story true? Would I actually say this sentence out loud?" 5:40 And if the answer is no, he forces the AI to revise it until the answer is yes. Still, the colleagues noticed. Right, but the feedback from colleagues isn't a sign that the system failed. 5:50 I mean, it certainly sounds like a failure of the ghostwriter to sound like the author. It is actually proof that the feedback loop is functioning correctly. The AI isn't just left to its own devices. 5:58 When those colleagues pointed out the unnatural phrasing, the author incorporated that feedback to refine the system instructions further. So it's an ongoing process. Yes. The human always remains the final filter. 6:09 You calibrate the machine over time. 6:11 Okay, I understand the theory of calibration, but let's look at the actual facts being presented here, because an awkward idiom is one thing, but a fabricated technical detail is entirely another. 6:21 Oh, sure, hallucinations are a real issue. Right, and this is where the stakes get much higher for a manager. If a phrase is slightly off, a reader might just find it quaint. 6:30 But if a technical fact is wrong, say the AI hallucinates a server limit or invents a software capability, a company could experience an outage or a client might be completely misled. 6:40 And that is exactly why the research and fact-checking protocol in this workflow is so remarkably rigorous. 6:46 To prevent the AI from making things up, it is strictly instructed to check technical facts against official documentation PDFs. PDFs provided by the author, right? Yes, provided directly by the author. 6:57 It is grounded in truth, not just guessing based on its, you know, its broad training data. And furthermore, if the AI searches the internet to fill in a blank, the author manually verifies the source. Manually. Right. 7:09 He does this specifically to ensure the AI isn't pulling inaccurate information from some random, unverified third-party blog. It has to point to official documentation. 7:19 But think about what you are actually asking a busy professional to do here. 7:23 Relying on an AI for this kind of technical verification is exactly like having a lightning-fast research assistant who sometimes lies with absolute confidence. Well, I wouldn't say lies. But it fabricates. 7:35 Yes, this assistant is incredibly fast. They can instantly find the exact name of a buried setting or pull up a specific usage limit that just changed last month. That speed is intoxicating. It is a massive time saver. 7:47 But is it? Because you know this assistant has a history of fabricating details, you still have to meticulously grade their homework. You have to trace the AI steps back to the official PDF. 7:57 You have to click and read the internet links. Right. So I have to ask, is the time you saved on the initial drafting process just being lost on the back-end fact-checking? 8:06 If I have to open the official PDF anyway to make sure the AI didn't hallucinate the limits, why didn't I just write the document myself? That is a very fair question. 8:14 But I think you're looking at the time spent fact-checking as a one-to-one penalty rather than recognizing the massive time differential in how we retrieve information. What do you mean? 8:23 Well, the human mind is not optimized for remembering the exact data limit changed in a dense technical manual. The AI is. 8:30 When the AI retrieves that detail and points you exactly to, say, paragraph four on page twenty-two of the PDF, verifying it takes a fraction of the time it would take for you to hunt it down from scratch. 8:40 But you still have to read paragraph four. Yes, but you are shifting from the role of a data miner to the role of an editor and approver. 8:46 It's the difference between asking someone, "Did this specific limit change on Tuesday?" And having to ask, "When did this limit change across all our documentation?" Hmm. 8:54 You are checking what matters rather than exhausting your mental energy just finding it. 8:58 Okay, I see the shift in roles, but the only reason that back-end fact-checking doesn't become a second full-time job is because of the rigid infrastructure holding the system together. The technical setup. 9:09 Yes, and this is where I think the blueprint completely alienates the average non-tech professional. We are no longer just talking about typing a prompt into a chat window or uploading a PDF. Right. It goes deeper. 9:21 Much deeper. The author's workflow utilizes a private GitHub repository, which for our listeners, we can explain GitHub as a digital filing cabinet with a perfect permanent memory of every draft. 9:33 Every single word spoken in the raw recording is saved there, version by version, and compared against the final output. It's incredibly thorough. And it doesn't stop there. 9:42 The workflow also utilizes something called MCP, the Model Context Protocol. 9:47 Think of MCP as a highly advanced digital switchboard operator that plugs directly into different software applications, like the newsletter platform, Google Drive, LinkedIn, Instagram, and YouTube. 9:58 Which is an incredibly powerful way to handle the heavy, repetitive parts of the work. It is powerful, yes, but look at the sheer complexity of what is being required here. 10:06 A private GitHub repository with version control, tracking word counts ranging from eight hundred to three thousand words. 10:13 A Model Context Protocol switchboard seamlessly integrating multiple social media platforms and cloud storage. It sounds like a lot when you list it all at once. Because it is a lot. 10:22 This level of infrastructure is highly intimidating. Does a non-tech manager really need a GitHub repo and an MCP setup just to write a weekly status email or summarize a team meeting? Well, no, not for a simple email. 10:35 The technological overhead to set this up seems completely disproportionate to the needs of an everyday professional. You are asking them to build a factory just to make one sandwich. 10:45 Okay, I will concede that setting up a GitHub repository is a steep learning curve. 10:50 The first time you try to version control a document, it feels incredibly unnatural compared to just, you know, typing in a standard Word doc. It is intimidating. Very. 10:59 But I'm not convinced by the factory analogy because it assumes you have to build the entire skyscraper on day one. You don't have to adopt the whole system immediately. But the author does. 11:09 Yes, but as the author specifically notes, you learn this post by post. Your knowledge grows with every single attempt. 11:16 Let's look at the mechanism of that MCP switchboard you mentioned, because it's not just moving text around. Think of MCP not just as a switchboard, but as a smart universal adapter. A universal adapter. How so? 11:28 Well, instead of a human manually taking a long-form essay and trying to cram it into a short-form LinkedIn post, the MCP understands the shape of the destination. The AI isn't just blasting text onto a screen. 11:40 It is handling the small mechanical details that make a massive difference. Like what kinds of details? 11:45 We are talking about crafting the perfect subject line, writing engaging preview text, resizing images so they fit the specific dimensions of each social channel automatically, generating alt text for visually impaired readers who use screen readers. 11:58 Okay, the alt text is actually a great feature. Right? And creating short, easy-to-type links. It routes the data to fit the pipes of each specific platform without the human lifting a finger. 12:09 Those details certainly matter if you are a full-time content creator or publisher. They matter to everyone who wants their communication to be effective. 12:17 A single poorly written subject line or the wrong image format can drastically change how many people read your message, whether it's a global newsletter or an internal company update. That's true. 12:27 If a human has to manually format alt text, resize five images, and write four different preview summaries for every single piece of communication, they will eventually stop doing it because it is exhausting. Yeah. 12:37 The fatigue is real. Exactly. So by letting the MCP, our universal adapter, handle these purely mechanical tool steps, the human is entirely freed up to focus on the message itself. 12:49 Even the diagrams are generated from code in the author's brand colors, requiring only a final human approval. I will grant you that the automation of alt text and image sizing is a brilliant use of the technology. 13:00 Relieving that specific burden makes total sense, but I wanna circle back to the GitHub repository, our digital filing cabinet. Okay. Why do we need permanent memory for writing a post? 13:10 The author states that even the workflow itself, the how we work document, is written down in that repository, so the AI follows the exact same steps every time. Right, for consistency. 13:21 But this requires a level of systems thinking that is fundamentally alien to how most people express themselves. Most people sit down at a keyboard and just write. 13:29 Asking them to engineer a repeatable, documented version control pipeline just to share their thoughts, it feels like we are turning human expression into an assembly line. 13:38 I'd argue we are turning the distribution into an assembly line precisely so the expression can remain deeply human. And that permanent memory in GitHub, it's vital because AI can change things unpredictably. 13:49 Unpredictably is a good word for it. Right. So if you have a perfect memory of every draft, you can revert instantly when the AI accidentally deletes a great paragraph or hallucinates a fact. 14:00 It gives you the safety net to experiment. I suppose a safety net is necessary. It is. Remember the core philosophy of the post. A five-minute read has 30 years of experience behind it. The stories are real. 14:13 They come from actual customer questions, from moments when a colleague simply messages, "Hi," and a complex problem unfolds. Yeah. AI was not in those rooms. 14:22 It doesn't have the memory of that specific client meeting 10 years ago, and it certainly doesn't know which specific story is most relevant to your team this week. And on that, we are in absolute agreement. 14:32 That is the golden rule of this entire framework. AI does not have your memories, it does not have your scars, and it does not have your intuition. Never. 14:40 Therefore, the genesis of the ideas and the absolute final decisions on what gets published must always remain strictly with a human. The human soul of the work simply cannot be outsourced. Exactly. 14:51 The human provides the 30 years. The AI provides the five minutes of reading polish. And I think we both see eye to eye on the non-negotiable role of human experience. We do. 15:00 Where we find nuance and where this really depends on the listener's own situation is how far down the automation rabbit hole you choose to go. Right. 15:07 Because if I am a manager listening to this, I am probably not setting up a GitHub repository and an MCP switchboard today. Probably not today, no. So what is the actual starting point? 15:18 Whether you just use AI to help format a cleaner version of an important email, or whether you go full throttle and automatically publish across five different platforms while logging every keystroke, that depends entirely on your personal technical comfort level. 15:32 And your workload. Yes. It depends heavily on how repetitive your specific weekly workload is. If you are doing highly bespoke, deeply emotional work, you might keep the AI at arm's length. 15:43 If you are aggregating weekly data reports, you might hand 90% of it to the switchboard operator. Which is exactly why the, the practical next step, what the author calls this Friday's tip, is so valuable. Why's that? 15:54 Because we wanna deliver the exact homework from the post for our listeners to try, and it weaves right into what you just said about finding, you know, your, your starting point. Mm-hmm. 16:03 First, write down your own workflow. Pick just one thing you do every week. Okay. So like, like a report, um, maybe it's a status email, maybe it's just this reoccurring presentation. 16:13 List every single step involved in that one task from start to finish. Then what you're gonna do is just you're gonna kinda go down the list, and then you're gonna just kinda mark it as, is that a me? Is that a tool? 16:24 And step two is where you have to practice strict discipline. Keep the me steps entirely for yourself, the ideas, the stories from your career, the decisions, and the final review must stay with you. 16:34 Do not outsource those. Right. But choose just one heavy repetitive step and hand that tool step to the AI this week. 16:42 And to your point about safety, choosing to hand off a formatting task like organizing bullet points or drafting alt text is a much safer first step than asking the AI to summarize a highly sensitive, controversial client meeting. 16:54 Definitely. Let it do the mechanical heavy lifting and then strictly review the result. And the final step, step three, ask two people for honest feedback on your output and change one thing based on what they say. 17:06 You don't ask for feedback to feel good. You ask for it to get better. Yeah. Incorporate it into your routine, just as the author did when his colleagues pointed out the nuances of his English phrasing. 17:14 You calibrate the machine. It is a measured, safe way to test the waters without handing the keys to your career over to an algorithm. You keep the wisdom, you outsource the typing. 17:24 As we close today, think about the tools you use every day. 17:28 Are you still developing the film by hand in a dark room, or are you ready to let the camera handle the focus so you can spend your valuable time simply looking for the perfect shot? 17:38 A fascinating question to leave them with. You can read the full post free at intechornot.com
