YT YouTube2Text

YouTube2Text → Transcripts → Fallo UNA VOLTA. Ogni AI lavorerà meglio per te

Fallo UNA VOLTA. Ogni AI lavorerà meglio per te — Transcript

by GabrySolution · 2,974 words · 442 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00There’s a cost you pay every single
0:02time you open an AI chat, whether it's
0:04ChatGPT, Gemini, Claude, or any other.
0:07You don't see it on your bank statement
0:08, yet you pay it religiously. It works
0:10like this: new conversation. And you
0:12start back at square one. Who are you?
0:14What does your company sell? How do you
0:15want your texts to sound? Who are you
0:17addressing? It gives you back something
0:19decent, but generic, and you burn
0:21minutes fixing everything. And tomorrow
0:23, tomorrow you start all over again.
0:25And sure, the platform you use today
0:27might remember a thing or two, but if
0:28you have to switch platforms,
0:29everything you taught it stays locked
0:31in there, and you find yourself back at
0:32the starting point. The uncomfortable
0:35truth is this. There is no prompt that
0:38can fix this, and there isn't a tool
0:40that can fix it either, because the
0:42flaw isn't in how you write your
0:44requests or in the platform you chose.
0:47It lies in the fact that every session
0:49starts without knowing much about you.
0:51In this video, I'll show you the system
0:53I built to close this loophole once and
0:56for all. I call it the context pyramid.
0:58You build it once, and from that moment
1:01on, any AI you use—today, in a year,
1:03or in five—will automatically get
1:05better at working for you. Don't
1:07believe me? Let's go take a look, as
1:09always. Everything is ready in the
1:10description below. Ready? Let's get
1:12started. Before we look at the pyramid,
1:14I want to give a precise name to the
1:16enemy: I call it the invisible tax, the
1:18price in time and mental energy you pay
1:21every time you start working with AI. A
1:23technology created to gift you hours,
1:25which instead devours them. And this
1:28tax has three distinct components.
1:31First, you're piecing tools together
1:32with constant copy-pasting—a prompt
1:34here, a screenshot there, an output
1:36from one platform dragged into another,
1:38crossing your fingers that the results
1:40hold up. Second, you have to introduce
1:43yourself in every conversation, and I
1:44swear, this drives me crazy. Every new
1:47chat is a blank slate; the AI doesn't
1:48know who you are or knows very little
1:50about you, it doesn't recognize your
1:52writing style, it has no idea who your
1:54ideal client is, and even when a
1:55platform has active memory, it doesn't
1:57remember much. The third is the
1:58collection of subscriptions. Five or
2:00six tools that cost you, and half of
2:02them you might not even use. And what's
2:03more, every week someone you follow
2:05announces the definitive tool, the
2:07brand-new one. The result is that you
2:09feel perpetually behind, and it starts
2:10all over again. So, that is the tax.
2:13And let me tell you, the problem even
2:14has an official name now: it's called
2:16Context Engineering, and it's
2:18considered the most valuable AI skill
2:20of 2026. To the point that Anthropic,
2:22the company that develops Claude, has
2:24published a dedicated guide on it. The
2:26basic concept is very simple: the
2:28quality of what AI gives back to you
2:30depends much more on the context you
2:32provide than on the magic formula you
2:34use in the prompt. And the answer to
2:36all of this, in my opinion, is the
2:38context pyramid. Before I show it to
2:40you, though, I need to clear up one
2:42word, because you hear the word "agent"
2:44everywhere these days, and it's almost
2:45always made more complicated than it
2:47needs to be. Anthropic defines it in a
2:49single sentence: "An agent is a model
2:51that uses tools within a loop." Hold on
2:53, let's translate that. The model is
2:55the AI itself—Claude, or GPT, Gemini,
2:58whichever you prefer. The tools are not
3:00the apps you're subscribed to; they are
3:02the concrete actions that the AI is
3:04capable of performing. Opening a
3:06document, filling out a spreadsheet,
3:07sending an email, doing an online
3:09search. Anything the AI can physically
3:12do is a tool, and the loop just means
3:14it doesn't stop at the first step; it
3:16performs an action, observes how it
3:18went, decides the next move, and
3:20restarts until the job is done. Now
3:24comes the part that really matters:
3:25what distinguishes an excellent agent
3:27from a mediocre model. It's almost
3:29never the model; it's the material you
3:31make available to it. What information
3:33can it consult, and how clear is the
3:34description of the task you are
3:35assigning to it? I'll explain it with
3:37an analogy. When a new person joins my
3:40team, I don't tell them "figure it out
3:41yourself"; I tell them about the
3:43company, I show them what good work
3:44looks like, I give them access to
3:46training materials, and I show them the
3:48management roadmap so they have a clear
3:49context. With AI, however, we do the
3:52exact opposite: we open the chat, and
3:54since we know they are super
3:55intelligent, we throw in two lines and
3:57then complain that the output is flat.
3:59And here is what I realized. After
4:00months of banging our heads against the
4:02wall, the solution is not yet another
4:04perfect prompt, nor yet another
4:06platform. The solution is to put your
4:08business down in writing, once and for
4:10all, in a format that is readable by
4:12any AI. This is the context pyramid.
4:15Put that way it sounds trivial, so let
4:17me explain why it works so well. Think
4:20of a well-made onboarding manual. When
4:23a new social media manager starts
4:24working with us, we give them a full
4:26onboarding with all the processes: what
4:27to do, what not to do, warnings,
4:29precautions, emergencies—everything.
4:31It’s not just a single block of words
4:33; it’s instructions. It has an index,
4:35tells you where to find what, and
4:36it’s the way we humans absorb
4:38information and, surprise, it’s
4:40exactly how artificial intelligence
4:42absorbs it best too. And here is the
4:44idea that completely shifted my
4:45perspective. You don't need to build
4:47agents; you need to build the material
4:49for agents to read. The agent is simply
4:52the AI that you point at your pyramid,
4:54asking it to do something. Today you
4:57have Claude read it, so your agent is
4:58Claude. Tomorrow there is GPT, so your
5:01agent is GPT. The AI is the
5:03interchangeable part, the pyramid is
5:04the piece that remains, and if you
5:06think it’s trivial, look at what the
5:07whole sector is doing. Developers all
5:11over the world, in countless projects,
5:13have adopted a standard tool called
5:15agents.md, a simple text file put
5:17inside the project that explains to the
5:20AI how to work on it. Today, even the
5:23Linux Foundation is supporting it. And
5:25Anthropic has launched "skills,"
5:27folders of instructions that Claude
5:29goes to read on its own when it needs
5:31them. A format so well-conceived that
5:33even OpenAI copied it exactly for GPTs.
5:37The message is clear: the future of AI
5:39isn't in prompts, it’s in
5:40well-written files. Now I’ll show you
5:43what it looks like, because the
5:45structure is disarmingly simple. Two
5:47levels, that's it. At the top of the
5:49pyramid is a folder I call "My Company.
5:51" Your business, written down once and
5:53for all. Inside, there are four files.
5:56The first is "Who I Am": name, what you
5:58do, that introductory sentence you’d
6:00use with a stranger—your so-called
6:02claim. The second is "My Business."
6:04What do you sell? Who are you targeting
6:06? What makes your approach different
6:08from your competitors '? And so on. The
6:10third is your "Tone of Voice," and
6:12it’s the most valuable file of all.
6:14In here, you put your style, the tone
6:16you maintain, the expressions you
6:17repeat often, and the words you refuse
6:19to use. A practical tip on this: if you
6:22speak the way you write, you’re lucky
6:25; you can feed it transcripts of your
6:27videos or your best online content.
6:30I’ve been recording videos for a long
6:31time, so my spoken and written styles
6:33now coincide, but for most people,
6:34that’s not the case. In that case,
6:36you need real written examples: emails
6:38you’ve sent, social media content,
6:40anything at all. And if the two voices
6:42are very far apart, nothing stops you
6:44from creating two separate files, one
6:45for spoken and one for written. The
6:48fourth is "My Offers." What are you
6:50selling right now? With the relative
6:52prices. Here yes, precise but concise.
6:54The more synthetic, the better. And a
6:57small warning: you will almost never
6:59touch this top level, unless you launch
7:01a new offer; it will remain practically
7:02static for a very long time. Then there
7:05is the base of the pyramid, the second
7:07level. Here, every project in your
7:09business gets a dedicated folder. Your
7:13Instagram content goes in there,
7:14YouTube channel content in there, a
7:16client you manage has their own folder,
7:18and each of these project folders
7:20always contains the exact same setup.
7:23Five subfolders, never one more, never
7:25one less. Instructions, voice,
7:27references, examples, and diary. Let's
7:30see them one by one. Instructions is
7:32the project guide file. What is it
7:34about? Who is it for, what are the
7:36characteristics of a well-done result,
7:38what to avoid like the plague, and it's
7:40also the place where you tell the AI
7:42where to find the rest of the material.
7:44It works like the index of a book.
7:46Voice is only for that project. It has
7:49its own tone, different from your
7:51general voice. It becomes fundamental
7:53when you work with clients. Everyone
7:55has their own way of expressing
7:56themselves, their vocabulary, their
7:57style. References. It collects
8:00everything the AI should read when it
8:02starts working on the project. Research
8:04you've done, transcripts, internal
8:06procedures. And speaking of procedures,
8:08here is a shortcut for you. The next
8:10time you perform a process, record your
8:12screen while doing it, and then ask the
8:14AI to turn that recording into a
8:16written procedure. You see it within
8:18models; for example, on Claude from the
8:20desktop app, clicking the plus button
8:21will let you record, so it's a crazy
8:23little hack. Five minutes and the
8:25documentation is ready. Examples is the
8:27folder preferred by the AI. Models
8:29learn a lot from concrete cases of
8:31successful work. If the project is
8:33Instagram content, put your four or
8:34five best posts ever in there. And then
8:36there's the diary, the subfolder where
8:38you note how things are going. And this
8:40is where the system really lights up.
8:43Let's say Tuesday's post did very well.
8:45Open the diary, leave a couple of lines
8:47, and the AI has the statistical data
8:48from that moment on. It no longer works
8:50in the dark. You are learning week
8:53after week what really works for you.
8:56Well, now you know the structure: two
8:58levels and five subfolders, and perhaps
9:00you're already thinking you'll have to
9:02spend an hour creating folders by hand
9:04and typing texts into files one by one.
9:07Absolutely not, because you'll do it
9:09with Cloud Cowork or ChatGPT Team, and
9:11the great thing is that you'll ask it
9:13to build everything for you: folders,
9:15files, content, everything. And now, to
9:17prove it to you, we'll do it together.
9:19I use Cloud Cowork, but I repeat, you
9:21can also do it with ChatGPT Team.
9:22I’ll write exactly this prompt. Build
9:24my context pyramid. At the highest
9:27level, create a folder called "My
9:29Company" with four files inside: My
9:31Company, My Voice, My Offers. Then,
9:34create a project folder called "
9:36Instagram" and inside it five
9:37subfolders: Instructions, Voice,
9:39References, Examples, Journal. For now,
9:42I'll leave all the files empty. At this
9:44point, Cowork asks me which folder on
9:45my computer it should operate in; I
9:47point it out, give consent, and it goes
9:49. Or you can directly select a folder
9:51yourself before launching the prompt,
9:53so Cowork will only work in that
9:54location. Clearly, you can also include
9:57it in the prompt. Once that’s done,
10:00launch the prompt and voilà, you’ll
10:01have everything ready in less than 5
10:03minutes. My Company at the top with its
10:05four files and the Instagram folder
10:07with all the subfolders. And now you'll
10:09say, "But how do I fill it all?" Well,
10:12we'll do it without even writing a line
10:14because the structure is ready, but
10:15filling it manually takes a really long
10:17time because we need to be precise. At
10:20this point, I write a second prompt to
10:22Cloud. Now, let's fill the files in the
10:24"My Company" folder. I'll use a
10:26fictional business. Ask me questions,
10:29interview me, and when you have enough
10:31information, write the content directly
10:33for me. Now, at this point, Cowork asks
10:35me all the questions, and look, it goes
10:37and distributes everything into the
10:39folders and files in the right place. A
10:42quick tip: to give it more information,
10:44instead of typing, you can also
10:46activate voice mode and speak, so it
10:48automatically inserts everything. And
10:52if you already possess those documents,
10:53for example, presentations, briefs, or
10:55old company files, it's even easier:
10:57connect the folder where they are
10:58located, or move those files into the
11:00folder that Cowork has access to and
11:01tell it, "In the folder I linked, there
11:03are various documents about my business
11:05." Read them all and distribute them
11:07into the pyramid. What relates to the
11:09company goes into the four "My Company"
11:11files. What relates to content goes
11:13into the Instagram project subfolder.
11:15If any of these files are no longer
11:17valid or correct, let me know. This way
11:20, it will do everything. If you open
11:21the files now, you see, it's all
11:23perfect. If you want to simplify things
11:25further, you can tell it: "Okay, in the
11:27folder for my company, I'll give you
11:28all the information," and start right
11:30away with a very long voice memo. It
11:32will take all the information from the
11:34voice memo and then distribute it into
11:35the files. Done? You do it once, 10
11:37minutes, and you're set forever. Now,
11:39however, I have to prove to you that it
11:41really works, and I'll do it twice with
11:43some quick tests. I'll open a brand-new
11:45session so it doesn't remember anything
11:47, or by clicking in the top right, you
11:49can even activate incognito mode, both
11:51on ChatGPT and Claude. This way, it
11:53truly has no memory of anything. At
11:56this point, I connect Claude to the
11:58pyramid folder and give it a direct
12:00task without any introductions. Here is
12:02the prompt. Read all the files in the
12:04context pyramid from the first to the
12:06last, then propose three reel ideas
12:10that I would really like to read
12:12written in my voice, each with a
12:13one-line explanation. Quick note, if
12:16you look at the Claude sidebar, you can
12:18see the list of files it's consulting
12:20in real-time; it's literally studying.
12:24Look at the result: three reel ideas
12:25built around the concrete problems from
12:28our example earlier. I’ll repeat it
12:29because the whole point is right here.
12:31I didn't copy and paste anything; I
12:33explained who I am, what I do, and what
12:35tone I want. I pointed it at the
12:37pyramid, it read it, and nothing else
12:39was needed. But let's do a second test.
12:41The second test is truly incredible.
12:43Remember the phrase from before? The AI
12:44is the interchangeable part, the
12:46pyramid is the part that remains, and
12:48I'll prove it to you right now. Using
12:50the exact same pyramid that lives on my
12:52computer locally, I'll feed it to
12:54another artificial intelligence that
12:55has never seen these files. I open
12:58ChatGPT, activate work mode, or attach
13:00all the files directly to the chat. I
13:03don't write anything; I can just drag
13:05them in and go. Or if they were, for
13:07example, on Google Drive, I just
13:08connect Google Drive. At that point, I
13:10say, "Read the attached files and
13:12propose three reel ideas for Instagram
13:13in line with what you find in there."
13:16The same request made to Claude, more
13:18or less, and look at what comes out
13:20from ChatGPT instead. See? It's truly
13:24crazy. The pyramid stays, the AI
13:26changes. And that is exactly why I
13:28consider it the only investment in
13:30artificial intelligence that never
13:31loses value, but actually increases
13:33because the more you do, the more your
13:35project expands, and the more
13:36information you put in there. Maybe at
13:39first, you have three viral pieces of
13:40content, then you'll have 10, then 20,
13:42so it becomes more and more precious. A
13:44model stops being the point; the best
13:46one of this month, the next, or in
13:47years, doesn't matter. I focus on the
13:50pyramid that already knows the whole
13:51context and knows everything about me.
13:53Think about it, some of the tools we
13:55use today will have changed name, owner
13:57, or otherwise in a few years; it's the
13:58nature of the market. But the pyramid
14:01you build now will still be there. It
14:03is increasingly important because when
14:05your business evolves, you update the
14:06file, and when you leave two lines in
14:07your journal, you are teaching it
14:09something. And it's not just my feeling
14:11, it's an academic study presented at
14:13ICLR 2026, one of the most important
14:16conferences in the field, which
14:18measured just this: treating context as
14:20a living document where you add what
14:23worked and remove what failed. It
14:26improved app performance by over 10%.
14:30Not only that, a small, free model,
14:31armed with a well-built context,
14:33matched the results of top-tier models
14:35that were left to their own devices. So
14:38imagine giving context to a much more
14:40powerful model instead. It is the
14:42winning pair; it is literally written
14:44in the data. So I have only one
14:46invitation: build it, test it, and work
14:48on it. Now, if this video was useful to
14:50you, leave a like, subscribe to the
14:52channel, and as always, happy solving.

About this transcript

This page contains the full transcript of Fallo UNA VOLTA. Ogni AI lavorerà meglio per te by GabrySolution, generated from the public captions YouTube serves with the video. The transcript has 2,974 words across 442 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.