# An AI that remembers everything about your company — build yours in 30 minutes > SecondBrain gives AI durable company memory: customers, decisions, procedures and lessons in a portable text-based system. Here is the architecture, real use cases and the complete setup bundle. Author: Fabrizio Romano · Published: May 14, 2026 · Reading time: 41 min read Source: https://futuriamarketing.com/en/blog/secondbrain-ai-azienda-30-minuti/ --- > *I built it for Futuria in three months. It is called SecondBrain. Today I am giving it to you: you can recreate it for your company in 30 minutes.* Five numbers, to start with something concrete. - **27** customers tracked as living entities — active customers, opportunities and partners. - **137** person profiles — customer contacts, decision-makers, prospects, partners and colleagues — each with an interaction history and notes on how they prefer to work. - **175** archived calls, each with a summary and links to the relevant customer, people and resulting decisions. - **62** written and applied operating procedures — from onboarding a new colleague to deploying a change on a customer’s website. - **0** person-hours spent each month keeping all of this up to date manually. I am not showing off. I am making a different point: a year ago, our working day began by opening eight tools — Google Drive, Gmail, Google Calendar, ClickUp, Futuria CRM, WhatsApp Web, ChatGPT and Slack — and trying to reconstruct from memory what had happened yesterday, with which customer. Now it begins by opening *one* place. That place already knows everything because it updated itself overnight. That place is called SecondBrain. It consists of text files. Humans and AI agents can both read it. It costs nothing in licence fees. And it can remain on your computer permanently, even if Anthropic, OpenAI, Google and Obsidian all close on the same day tomorrow. This article explains what it really is, why it works, how we built ours, how we set it up for Jacopo, Massimo and Farwa, and — at the end — provides a downloadable package containing everything you need to start your own. This is not a generic Obsidian tutorial. It condenses a month of real setups for real people into a bundle that you open with Claude Code or Codex and tell to “start”. An important note, because choosing the wrong tool here is very easy: **you need Claude Code or Codex**. Not ChatGPT, Gemini or the ordinary browser version of Claude. The reason is simple: the setup must read, create and edit files on your computer. A generic chat can discuss the prompt, but it cannot actually build the vault for you. If you are looking for five prompts that will revolutionise your life, you are in the wrong place. If, however, you have spent months feeling the frustration of company memory dissolving whenever you close a chat, stay. This article is for you. --- ## Summary 1. What a SecondBrain really is — and what it is not 2. The four components of an agent — and why the model is the least important 3. The structure: PARA, and the difference between brain and body 4. Three levels of memory — the pumpkin-recipe analogy 5. Four things that genuinely change in your working day 6. Three real setups: Jacopo, Massimo and Farwa 7. The routines that keep the vault alive 8. What does not work — and why you need guardrails 9. The package to download 10. A final thought about a Makita drill --- ## 1. What a SecondBrain really is — and what it is not Let us remove the ambiguity first, because “SecondBrain” is one of those terms everyone uses differently until nobody understands anyone else. A SecondBrain is **not** a well-organised Google Drive. It is not a folder containing 200 documents you swear you will read again. It is not a collection of linked Google Docs. Nor is it an Obsidian vault full of your chaotic notes from 2022. A SecondBrain is **the operating system for your mind**, made from text files organised according to strict rules. It gives a structured home to: - the work you have done; - current projects; - decisions and the reasons behind them; - the people you know and what you discussed; - things that did not work and why; - repeatable procedures; - credential metadata — not the secret values themselves. It uses a format that both you and an AI agent can read and apply. The real difference from Google Drive and similar systems is not cosmetic: **the knowledge is navigable by AI as though it were a personal Wikipedia**. Every file links directly to others, and the agent explores them as you would move from one Wikipedia page to the next. When you ask, *“What did we decide in March with that construction-sector customer?”*, it does not search the cloud hoping to find the right sentence. It opens the customer folder, reads its overview, follows the link to the decisions file, opens it, tells you what was decided and cites the line. The file remains readable to you. You can see *exactly* what the agent is reading. If it is wrong, you know where to correct it. If you change AI model tomorrow — moving from Claude to another — the vault works in exactly the same way because it is only text. The knowledge belongs to you, not to the provider. That combination — navigability, portability and everything being readable in plain text — is the difference between a real second brain and another application you will forget in six months. There is also a subtler change, one you only understand after living with the system for a few weeks: **the vault stops being a place where you put things. It becomes a place from which things emerge ready to use.** During the day, you produce material: a hastily written call note, a decision made on the spot, a doubt set aside, an important email you do not want to lose. All of it goes into one folder called `00\_Inbox/`. You do not classify it or move it yourself. Overnight, an automated procedure — the agent — reads it, understands what it concerns and moves it where it belongs. Notes from a customer call enter that customer’s folder. Advertising questions enter the advertising area. Strategic decisions go into the correct decisions file. The next morning, you open the vault and the inbox is empty. Yesterday’s knowledge is already in the right place, already linked and already retrievable. That is the promise. Now let us see how it actually works. --- ## 2. The four components of an agent — and why the model is the least important When people think “AI”, they think of the model: ChatGPT, Claude or Gemini. It is the part everyone sees. It is also the least important part. A useful AI agent consists of four things, in their true order of importance: 1. **Context** — the data it can access in real time; what it sees while answering. 2. **Memory** — how it stores and retrieves information over time; what it remembers about you between sessions. 3. **Tools** — what it can do. Not merely read, but act: open files, call an API or send an email. 4. **Intelligence** — the underlying model: Claude, GPT or anything else. It sounds like a technical detail. It is the most important point in this article, so I will repeat it. **An agent with good context and memory, even when powered by a mediocre model, will always outperform one with the best model in the world but no memory and no access to anything.** Always, without exception. In an entrepreneur’s language: you do not need the most expensive AI model. You need a well-built vault. I understand this well because customers come to us after making exactly the opposite mistake. One arrived after spending fourteen thousand euros on an “all-in-one AI” platform that *nobody on the team had ever really used*. Not the owner, either of the two salespeople or the marketer. It was a box running automatically while the real work continued to live in WhatsApp and email. The underlying AI model was excellent. The context it could access was zero. The memory it was building was zero. The result was a monthly bill paying a supplier for a service that had never actually started. AI by itself does not make a company work. It is the substrate that can make it work once the other three elements have been built on top. Therefore, when I say “agent” in this article, I do not mean a generic AI chat. I mean an operating environment such as **Claude Code** or **Codex**, opened on a local folder and permitted to read the bundle, create files, organise folders, check what already exists and stop before it risks overwriting anything. ChatGPT, Gemini or Claude in a browser can help you understand the idea. They are not the right tools for executing this setup. One practical consequence follows: **knowledge must not live inside the AI’s head. It must live in the system.** This is the principle of interchangeability, and it solves a problem everyone who uses AI for work faces every month. Every week brings a new model. Anthropic releases a more powerful Claude while OpenAI announces GPT-5. Google produces a faster Gemini. The following day, somebody says Codex is actually better for coding, Claude is preferable for writing and a different model is required for structured research. You try it, adapt and become comfortable — and two months later another model arrives that you are supposed to test. Meanwhile, every change of tool creates the same awkward doubt: *“This one now knows so much about me; I have spent months training it. If I change, do I start from zero?”* SecondBrain removes that problem. Knowledge does not live *in the tool’s memory*. It lives locally in the Markdown files in your vault. Tomorrow, you may move from Claude to Codex because a new Codex release does something you need. Open Codex on the vault and it can access exactly the same knowledge Claude had. No repeated onboarding and no starting again. The agent changes; the knowledge remains yours, in your files, on your computer. And when — inevitably — a model better than both appears, you open it on the vault and continue from where you left off. The model is the component that changes every three months. The vault remains for years. That is the first principle to retain: context and memory matter more than the model. The rest of the article explains how to build them. --- ## 3. The structure: PARA, and the difference between brain and body Now for the vault structure. It sounds dull. In reality, it is the point on which everything else stands or collapses. The vault has six folders at its root. Six — not seven and not five. It is PARA with one addition of our own: - `00\_Inbox/` — limbo. Everything that arrives and has not yet been classified. - `10\_Progetti-interni/` — internal projects with an end: things you eventually close. - `20\_Areas/` — continuing areas of responsibility: the agency, the product, personal finances or health. They have no deadline and remain indefinitely. - `30\_Resources/` — reusable knowledge: procedures, people who are not customers, credential metadata, cross-customer knowledge and templates. - `40\_Progetti-esterni/` — where most of the work lives: one folder for every active or prospective customer. - `90\_Archive/` — the archive. When a project closes, it moves here. It is never deleted, only moved. Six places and six functions. Before adding anything new, ask which of the six should contain it. If the answer is “I do not know”, do not create a seventh folder. Reformulate the question. An answer almost always exists; if it does not, the item belongs in `00\_Inbox/` until you understand it. Exceptions look harmless one at a time: a new folder here, a subfolder there, one misplaced file because “it is only one”. That is how vaults die. Three months later, you can no longer find anything and what looked organised has returned to being chaos dressed as order. There is also a fundamental distinction I want to name without turning it into an article within the article: **SecondBrain is only the brain. It also has a separate body.** The vault contains readable files — Markdown, text, decisions and knowledge. Everything else — source video, Photoshop files, code repositories, enormous datasets and renders — lives in another folder alongside the vault that we call the Workspace. With that distinction made, let us return to the brain. Every vault folder has a file named `\_overview.md`. This is the local map: it declares what belongs in the folder, what *does not* belong there and links to its main contents. Without `\_overview.md`, the folder does not exist for the agent. Literally: the agent always starts from the overview and does not enter if none exists. This is what keeps a vault navigable even after it grows to a thousand files. The agent does not become lost because it does not wander: it moves from overview to overview, each of which maps where to go next. Seen from a distance, this structure no longer resembles an organised folder. It resembles a graph: customers, areas, calls, people, decisions and procedures referring to one another. ![Graph view of the SecondBrain vault in Obsidian, showing linked areas, projects, resources and notes](https://assets.cdn.filesafe.space/jOFcdS1QIxGsOiKRyQti/media/219d0cd0-6bf7-483d-912c-117347dfeee4.png) The close-up view makes the point more clearly: the value is not “having many files”. It is having files that know where to point. When the agent enters the vault, it does not search blindly through a heap of documents; it follows explicit relationships. ![Detailed Obsidian graph of SecondBrain, with linked operational notes, people, procedures and areas](https://assets.cdn.filesafe.space/jOFcdS1QIxGsOiKRyQti/media/8ac80cae-0565-4226-b925-6ca055d43a5b.png) The other important item is the `CLAUDE.md` file at the vault root. It is the global operating contract. It tells the agent: you are in my vault, my name is Fabrizio, this is how I work, these are my rules and areas, and this is where things belong. It is the first file the agent reads whenever you open a session. Without it, the agent is working blind. That is the skeleton: six root folders, one overview in each folder and one global contract at the root. It is 80% of what you need to begin. --- ## 4. Three levels of memory — the pumpkin-recipe analogy Here is a metaphor I have used since I began explaining this system. Its usefulness was confirmed when I used it with Jacopo during his setup call in May. Think of a good chef who has worked in a kitchen for twenty years. That chef has broad **competence**. They know how to cook: when to add salt, how to finish with cold butter, that pasta should be drained a minute early. They do not look these things up every time; the knowledge is part of them. Then they have **recipes in their head**: the ones they prepare most often and that come naturally. They may not remember every exact proportion, but they retain the logic, structure and principal steps. They retrieve them quickly when needed. Finally, they have **a book of pumpkin recipes on a kitchen shelf**. It contains specifics: the number of grams for four people, the temperature and the cooking time. They do not memorise those details. They know *the information exists*, know *where to find it*, and look it up only when preparing that specific dish. An AI agent’s memory in your vault works in exactly the same way, on three levels: **Level 1 — Always active.** This is the root `CLAUDE.md`: global rules, identity, vault structure and tone. The agent reads it at the beginning of every session. It is the chef’s general ability to cook: rules that are always present and not open to debate. **Level 2 — Quickly retrievable.** These are the `\_overview.md` files in each area, together with the `memory.md` and `decisions.md` files for individual customers or areas. The agent reads them when entering that part of the vault. *“Am I entering a customer project? I open the customer overview, inspect its memory and check for relevant decisions.”* This is the recipe in the chef’s head: retrieved when required, but not kept loaded continuously. **Level 3 — On demand.** These are specific files: the transcript of a call from three months ago, a technical document about a particular integration, or version 7 of the code for a gift option that failed because of Iubenda. The agent retrieves them only when they are needed for that specific answer. This is the book of pumpkin recipes: opened when cooking pumpkin, and only then. This three-level structure solves the most basic and recurrent problem with AI agents: **they forget**. Most agent “memory” is merely the context window of the current session, so every new conversation begins from zero. Not with these three levels. The agent always starts with `CLAUDE.md` loaded, so it knows who you are, how you work and what your company does. When you say, “Let us work on Farway,” it opens Farway’s overview, which already contains historical decisions, people, tone of voice and the technical stack. In three files, you are providing what a new employee might receive across three days of onboarding. If the agent needs a detail — the exact checkout version, the specification for a CSV field or the transcript of a call with the founder — it knows precisely which file to read. It does not know everything. It knows where to find everything. That is the difference. --- ## 5. Four things that genuinely change in your working day You may now understand the theory while still wondering: “Fine, but what actually changes for me on Monday morning?” You already know AI can help solve many problems. You have probably used it for months. Yet you also know that every time, you must explain everything again: who you are, what you do, which customer you mean and what the context was. Even after explaining it for the third time in the same week, it never quite understands because the next session begins with everything forgotten again. SecondBrain removes that repetition. The agent already has the context, all the time. In practice, work that now takes half a day can take fifteen minutes. Not because AI replaces you, but because it stops making you combine *in your head* information that currently lives in five different places. Even apparently simple tasks — and this is the part people underestimate — usually require information from three or four separate sources. Here are four real examples of work we do now that used to be considerably more painful. ### Case 1 — You are about to join a call with a customer you have not spoken to for weeks Your call begins in five minutes. You have not spoken with this customer for six weeks. You glance at the calendar, see the name and realise that you do not clearly remember your last conversation. Twenty other customers, three sales calls, a technical incident, two trade fairs and a weekend during which you hoped to switch off have passed in between. You open the vault and tell the agent: *“I am about to join a call with \[customer\]. Brief me in 30 seconds.”* In 30 real seconds, the agent does four things together: - It opens the person profile — who they are, their role, whether they are the decision-maker or contact, how they prefer to be treated, words to avoid and personal quirks. One person is sensitive about pricing; another always wants the figures before the concept. - It reads the overview of the company they represent — sector, size, how long you have worked together and what is currently in progress. - It reviews the last three calls — subjects discussed, decisions made and outstanding work. - It checks `decisions.md` for any recent decision they may raise. It returns a brief: *“They probably want to discuss X. The latest decision on this was made on 18 April and was Y. Avoid renaming the product a ‘system’ — they always call it a ‘platform’ and care about the distinction. They have a tax deadline at the end of the month and are probably under pressure.”* You enter the call already thinking about the right conversation. No “Excuse me while I find the email”. No pretending to remember. No hurriedly taking notes about things they already told you. Consider how many versions of this scene occur during your day. For someone managing 20 to 30 active customers, it is half their appointments. Without the vault, those 30 seconds become three minutes of searching plus fifteen minutes of a less productive conversation. With it, they remain 30 seconds and the conversation begins at the right point. ### Case 2 — Positioning your product on Google without being an SEO expert Until a year ago, understanding how to position a business in search meant calling a senior SEO consultant, paying three or four thousand euros, receiving a PDF three weeks later and then being unsure what to do with it. With the vault, it can now be a day’s work. Consider [Farway Milano](https://farwaymilano.com), a premium direct-to-consumer children’s clothing brand from Milan founded by Farwa Zulfiqar. On 28 April, we produced updated keyword research for the company in half an hour — literally, not figuratively. We had not discovered a magic tool. The vault had already contained for weeks the things an external SEO consultant would first have needed us to explain: - the brand’s **design system**: colour palette, fonts and “smart luxury, anti-fast-fashion” positioning; - its established **tone of voice**: direct response, anti-luxury and authentic storytelling, derived from weeks of email-marketing work; - the **real products** in the catalogue — 72 products with variants already loaded on the website; - the brand’s **history** — relaunched in March 2026 after losing €50,000 when a Brera retailer failed; - the **welcome-series emails** being tested, and therefore the angles already working with the audience; - the **strategic decisions** made during calls with the founder over the previous two months. The agent read all of this, combined it with external market data — search volumes, intent and competitors — and produced research that genuinely spoke about Farway. It was not a generic PDF whose conclusion was “target children’s clothing”. It mapped keywords to funnel stages, associated the right terms with each angle already validated in the welcome-series copy and excluded searches that should *not* be pursued: terms that were too broad or too close to traditional luxury for Farway’s audience. A senior SEO consultant might have needed four days to produce the same work: first two calls to understand the brand, then the design system, the products and the tone of voice. The vault already held all of it. The operational point is simple: **a task that looks straightforward — keyword research — actually requires brand, products, positioning, market, tone of voice, previous experiments and external data to be combined**. Without one home for that material, you recreate the combination by hand every time. With the vault, the agent can assemble it in half an hour. Meanwhile Farwa, the founder and CEO, can continue doing her work — selling children’s clothing — instead of spending two days being interviewed by an external consultant. ### Case 3 — Changing something on a customer website you built three months ago This happens to us two or three times a week and, without the vault, is a small nightmare. The customer writes: *“Hello, I need to change that thing on the checkout. Do you remember it?”* You do not. Three months earlier, you had created nine different versions of the checkout code. Not because you were incompetent, but because a cookie-compliance system — Iubenda — intercepted every script that touched cookies and blocked it nine times. The tenth version finally worked: the script was an external file served from the same domain; cookie-related patterns were obfuscated in the source; and the DOM was inserted as a sibling rather than a child of the WooCommerce React subtree so a re-render would not remove it whenever the customer changed payment method. Three months later, you remember none of that. Without the vault, you rediscover it by repeating the mistakes. Three days disappear. With the vault, this happens instead. You open Claude Code or Codex and write: *“I need to change Farway’s gift option at checkout. Remind me what we did and why.”* The agent opens `40\_Progetti-esterni/farway/lessons.md`, finds the gift-option entries and summarises them in 60 seconds: the DOM injection, the Iubenda workaround, the official `extensionCartUpdate` pattern adopted for version 10, and the fact that the snippet scope *must* be `global`, not `front-end`. It cites the canonical code: `snippet20\_gift-option.php`, ID 39 in the production database, currently `active=0`. It also tells you where the database credential metadata lives — `30\_Resources/Credentials/farway--mysql-db.md` — how the SSH and MySQL deployment works, using an existing Python script, and reminds you that the SSH path on the server differs from the SFTP path. Confusing the two already cost an hour last time. There is an important point beneath the surface: the agent did not simply “know” all of this. It went to read it. It retrieved material from `lessons.md`, `decisions.md`, the credential file and the worklog — four different sources combined into one answer. You sit down and complete the change in two hours rather than three days. The difference is not that AI wrote the code for you; it did not. The difference is that AI provided in three minutes the context that would have cost half a day to reconstruct mentally. ### Case 4 — Preparing a demo, icebreaker or sales call that begins in the right place The final example may be the most valuable because it directly affects sales. Before AI, preparing a Futuria CRM demo for an unfamiliar company looked like this: the salesperson opened LinkedIn, read a few things, opened the company website, opened Atoka for financial information, opened the prospect’s email, took notes on a sheet and joined the call five minutes early hoping to remember it all. Preparation took 40 minutes per call, was performed differently by different people and produced wildly inconsistent results. The first 40 minutes of the call were then routinely spent asking the basic questions again — *“Tell us who you are, how long you have operated, how many people you employ and what you do”* — because nobody could reach specifics without first rebuilding the picture. Now, before the call, the agent prepares a standardised **company profile** for the prospect. It draws from: - **Atoka or Cerved** for registered details, size and revenue where available; - the prospect’s **website** for positioning and language; - the decision-maker’s **LinkedIn** profile for role, professional history and published content; - any **emails** exchanged before the call; - our team’s **previous calls** with that person or somebody at the same company, if any. The agent uses a standard format and **declares its sources**. Better still, when it cannot find a fact, it says so explicitly instead of inventing it. Consider a real example: Market Color, a long-established microbusiness in Rubiera, Reggio Emilia, founded in 1987 with four employees. When we prepared our first call with founder Alessandro Iaccheri, the agent found the employee count, history and size. It did not find revenue because no reliable public source was available. The profile therefore said: *“Revenue remains `unconfirmed`. Until a stronger source or official company report becomes available, no revenue figure should be used in operational or strategic documents.”* That is the difference between useful and harmful AI. Harmful AI would have invented a plausible number. Ours wrote, “I do not know; do not write a figure anywhere.” It was instructed to behave that way. This is a guardrail, which I will discuss shortly. In the call, instead of asking Alessandro to spend 35 minutes telling us who they were, he found us already prepared. He felt taken seriously. We recovered half an hour, and the conversation began in the right — strategic — place rather than at phase zero. Again, under the bonnet, that company profile combines six or seven sources. A capable person can build one in 40 minutes. The agent does it in five and does so more consistently because it never forgets a source. --- Four cases with one underlying theme: **apparently simple daily tasks actually require information from different places to be combined**. The value of SecondBrain is not that “it does one thing”. It holds things together, and when it combines four or five of them, it gives you hours of your life back. Now let us look at what happened when we built the system for other people. --- ## 6. Three real setups: Jacopo, Massimo and Farwa We first built SecondBrain for Futuria because we needed it ourselves. Once it worked, we began giving it to others. Over one month, three different people brought three different situations. Each setup taught us something that became part of the bundle I distribute today. I will describe them chronologically because the evolution of the system becomes clear. ### Jacopo — the founder who produced the canonical template **Jacopo Viola** is a founder managing several projects in parallel. His principal focus is Videogo, a video-sector startup he intends to make his core business. Meanwhile, other operational commitments keep him busy as Videogo grows. Multiple projects, multiple contexts and repeated AI sessions beginning from zero: a lifetime of knowledge living inside chats and dissolving with them. On 5 May, we spent an hour on a call. The result was not a vault hurriedly built during the call but a map of *what Jacopo’s vault would look like* if he built it himself. Afterwards, I sent him a written deliverable — a single Markdown file of roughly ten pages — describing his specific structure, areas, routing rules and adapted `CLAUDE.md`. He opens Claude Code, loads the file, writes *“Let us begin the vault onboarding”*, and the agent continues from there. That call and deliverable produced something I had not expected: **the bundle’s canonical template**. Setting up Jacopo’s system forced me to formalise things that had seemed obvious to me but were not documented anywhere — the distinction between vault and Workspace, the exact meaning of `\_overview.md`, the hierarchy of rules and the three maintenance routines. Those materials now live in Futuria’s vault as reusable operating procedures and have become the package I distribute. The most important lesson from Jacopo is very simple: **to teach something, you must first codify it**. Building the bundle required a real setup for a real person. It would never have emerged from my head in the abstract, even if I had tried for a month. ### Massimo — the broker who taught me to split the setup into smaller parts **Massimo Ciprandi** is an insurance broker and founder of Aegis Intermedia, managing 1,300 customers and working with 26 appointed insurance companies. At the same time, he is building Business Roundtable, an entrepreneur network that he is franchising across other Italian cities: monthly events with about 80 participants and chapters opening in multiple locations. It is an evolution of the familiar BNI model at a much lower membership cost. When I proposed building his SecondBrain, he immediately gave me an important correction: *“Fabrizio, I never have a full uninterrupted hour. Never. I cannot tell you when I will.”* The standard setup used with Jacopo was a continuous one- or two-hour session from beginning to end. It could not work for Massimo. So we experimented by **splitting it into two macro phases**, separated by as much time as required. - **Macro phase A** — 15 real minutes. The agent assumes sensible defaults — four predefined areas based on his profile: Aegis, Business Roundtable, Group and Personal — confirms them together and scaffolds the system. At the end, Massimo has a scaffolded vault, the Obsidian plugins installed and one populated example `\_overview.md`. He can close it and return to work. - **Macro phase B** — whenever he has 30 to 60 minutes of genuine concentration. The agent interviews him one area at a time, using three focused questions for each, populates memory, decisions and worklog files where real data emerges, and configures the overnight routines for his actual stack: Zoho Mail, WhatsApp Business and the ZO system he uses for the Business Roundtable protocol. Massimo entered the system without blocking two hours that did not exist in his calendar. The vault grew in stages calibrated to the way he actually works. The lesson became canonical: **the bundle adapts to the person, not the other way round**. Now, before delivering the package, I assess who the recipient is, how they work and how much uninterrupted time they genuinely have, then calibrate the flow. Busy founders like Massimo receive the two-macro-phase variation. People with a continuous hour receive the linear setup. More complex cases receive an even more segmented version. ### Farwa — the customer who is also a user of the system **Farwa Zulfiqar** is the founder of [Farway Milano](https://farwaymilano.com), the children’s clothing brand mentioned earlier. Hers is the most interesting case because it shows what happens when SecondBrain enters the relationship between an agency and its customer, rather than only the mind of the individual who builds it. Everything we do with Farwa lives in one place: the WooCommerce website and its custom PHP snippets, the March relaunch after the €50,000 loss caused by a retailer’s failure, the six-email welcome series in Futuria CRM, the design system and the direct-response, anti-luxury tone of voice. It is organised so that anybody on Futuria’s team can enter and recover the thread in five minutes. That is what allows keyword research to emerge in half an hour, a call to begin in the right place and a technical checkout change not to consume three days of “getting back up to speed”. Here is the important point: **SecondBrain does not only help the person who builds it. It helps the people inside the organisations that use it serve that person better.** The customer receives the benefits without necessarily knowing the system exists. They simply see an agency that answers more quickly, remembers everything and completes the work correctly at the first attempt instead of after three rounds of email. This is also why distributing the bundle across an agency or sales team has a cascade effect far greater than adopting it individually. One person with SecondBrain serves 20 customers better. Five people using it together through a shared vault multiply that effect. --- Three people and three versions of the same system. Jacopo produced the template. Massimo demonstrated that the template must be divided around real time constraints. Farwa confirmed that the vault serves not only you, but the people who work with you. --- ## 7. The routines that keep the vault alive I promised not to turn this into a technical lecture. I will keep that promise. In four paragraphs, here is what runs over Futuria’s vault each night and why the system would decay within months without these three routines. **One.** A routine named `nightly-ingest` starts every night at 22:01. It reviews the day’s recorded and transcribed calls, relevant emails and notes dropped into the inbox, then puts everything in its correct place. A customer note enters that customer’s folder. A strategic decision enters the correct decisions file. You open the vault in the morning and the inbox is empty. **Two.** A routine named `hygiene` runs three times a week — Monday, Wednesday and Friday at 09:01. It compresses redundant worklog entries, identifies archive candidates and reports structural anomalies. It is the maintenance work you might do at the weekend if you felt like it. Spoiler: you do not. **Three.** A routine named `weekly-profile-update` runs early on Monday morning. It updates colleagues’ person profiles from the previous week’s calls. If a one-to-one reveals a new working preference, a difficulty or a change of role, that person’s profile contains the update by Monday morning. All these routines share a property that is rarely discussed outside technical work but should be used everywhere: they are **idempotent**. You can run them ten consecutive times and receive the same result as running them once. No duplicates, repeated notes or uncertainty about whether something has already been done. It is the property that separates automation that helps from automation you must supervise. An important caveat: **these routines are not installed immediately**. When you use the package to set up SecondBrain, they are available as written procedures but are not active. You enable them when the vault contains real material — perhaps during the third or fourth week — because running them against an empty vault achieves nothing. --- ## 8. What does not work — and why you need guardrails I promised an honest article, so here are the things that must be managed properly if the system is not to turn against you: three painful lessons not to skip. **One: AI without guardrails is a time bomb.** We once generated a visual asset for an internal post, not a customer. AI placed a non-existent logo on a real company. We caught it in time and did not publish it. Since then, the vault has contained an `asset-generation-policy` that explicitly forbids the agent from inventing logos, brand names, data, revenue figures or campaign claims. When a visual asset needs a customer logo, the agent prepares the layout with a reserved blank space. I add the logo manually — or the agent retrieves it from the customer’s real brand folder if available. It sounds trivial. Without that rule, over six months we might have produced one inaccurate asset every month. Eventually one would have been published. **Two: no communication leaves the agent without human approval.** We once sent an empty email to a customer — a real incident caused by a faulty workflow rather than AI, but the principle is identical. The agent prepares the draft; a human clicks *Send*. Always. The fact that AI can technically send email autonomously does not mean it should. We are inflexible on this point. **Three: customer-service AI that exists only because other companies have it is theatre.** I have seen companies put AI agents into customer chat because they were embarrassed not to have them. There was no underlying process, memory or guardrail. The result is always the same: customers become angry at a stupid bot, employees must intervene anyway and nothing is genuinely saved. The real test for any AI you want to add to a business is: *what would change operationally without it?* If the answer is “nothing important”, you do not need the AI. It is merely a cost. Above all: **AI is not a substitute for a process**. If your company process is crooked, automation only accelerates it. The crash arrives sooner and hits harder. --- ## 9. Download the package Here is the part you have been waiting for. Everything described so far — the PARA structure, three levels of memory, onboarding agent, overnight routines, guardrails and personal calibration — is packaged as a series of Markdown files to use inside **Claude Code** or **Codex**. I want to be deliberately direct: **do not paste it into ChatGPT, Gemini, Claude.ai or another ordinary AI chat**. Not because those tools are poor, but because you do not merely need an answer. You need an agent that works on local files: it must download the bundle, read instructions, create folders, write Markdown, check whether Obsidian and Python are installed, and stop if it finds a conflict. In summary, the package contains: - the complete **conceptual framework** — the agent’s four components, the access hierarchy, the three levels of memory and the canonical principles: what you need to understand before starting; - the **vault architecture** — PARA, the separation between brain and body, and routing rules: what you need while building; - the **`CLAUDE.md` starter** — the operating-contract template to populate at the vault root, which the agent will read at the beginning of every session; - the **onboarding agent** — the SecondBrain Assistant persona and its phased protocol, guiding you from the first `mkdir` command to the first populated `\_overview.md`. You load the file, write `start` — or “Let us begin the vault onboarding” — and it continues from there; - the **maintenance routines** — `nightly-ingest`, `hygiene`, `weekly-profile-update` and the `vault-lint` pre-flight check: abstract procedures to install once the vault contains real material; - a **README** explaining how to use the bundle and what to expect. Here is the route map before you begin: 1. Open an empty local folder — or one dedicated to your new SecondBrain. 2. Open **Claude Code** or **Codex** in that folder. 3. Paste the prompt below. 4. The agent downloads the bundle from GitHub and reads its files in the correct order. 5. It checks the environment, Obsidian, Python and any existing vault. 6. It conducts a short interview to understand who you are, what you manage and how much time you have. 7. It creates or adapts the vault structure: folders, `\_overview.md`, `CLAUDE.md`, routines and `vault-lint.py`. 8. It closes macro phase A with a verification, then proposes macro phase B, during which the system is populated with real customers, areas, people and decisions. What you need beforehand: - **Claude Code** or **Codex** installed on your computer — one is enough. - **30 minutes** of genuine concentration, once. That is enough to scaffold the vault, activate the agent, define the areas and populate `CLAUDE.md` — a system that is already operational and already yours. You can populate the areas with customers, decisions and people in ten-minute stages over the following days, whenever a gap opens between calls. You do not need to block half a day. Installing Obsidian, scaffolding folders, enabling plugins and configuring `CLAUDE.md` are all part of the guided setup. You do not need to prepare or download anything manually. The bundle lives in a public GitHub repository, and the agent clones it automatically at the beginning of the first session. The experience is as follows. Open Claude Code or Codex in a clean session, select the most advanced model available and enable its highest reasoning level. Paste the single prompt below and submit it. The agent clones the bundle, reads the files in the correct order, assumes the role of “Futuria SecondBrain Assistant” and begins macro phase A of the onboarding. It asks a handful of focused questions — who you are, which computer you use and what you manage in your working life — then proposes a calibrated structure. Once you confirm it, the agent starts building. By the end of the session, you have a vault scaffolded on your drive, defined areas, a populated example `\_overview.md` and the complete structure required to add real knowledge over the following days. If your diary is as crowded as Massimo’s and you do not have 30 continuous minutes, the agent can divide the setup into two even shorter stages; that variation became canonical precisely for cases like his. One final operational reminder before you copy it: **paste this prompt only into Claude Code or Codex opened in the folder where you want to work**. If you paste it into ChatGPT, Gemini or ordinary Claude, the most you will receive is an elegant explanation of the setup. The vault will remain unbuilt. Marvellous rhetoric; dreadful for Monday morning. Copy and paste this prompt — nothing else: ``` Clone or update the repository https://github.com/FuturiaMarketing/secondbrain-bundle inside a `bundle/` subfolder of the current folder. If `bundle/` does not exist, clone it there. If `bundle/` already exists and is the same Git repository, update it to the latest version. If `bundle/` exists but is not that repository, do not delete anything: stop and ask me how to proceed. After downloading it, read these files in order: 1. `bundle/README.md` 2. `bundle/BOOTSTRAP.md` 3. `bundle/.claude/skills/secondbrain-assistant/SKILL.md` 4. `bundle/01-framework-e-principi.md` 5. `bundle/02-architettura-vault-e-workspace.md` 6. `bundle/03-claude-md-starter.md` 7. `bundle/04-agente-secondbrain-assistant.md` 8. `bundle/05-stack-discovery.md` 9. every Markdown file inside `bundle/routines/` 10. `bundle/meta/scripts/vault-lint.py` Then assume the role of “Futuria SecondBrain Assistant” as described in the bundle. Your objective is to guide me until I have a personal SecondBrain that is installed, configured and maintainable. Begin with macro phase A of the onboarding and follow the process defined in the bundle: - detect the environment automatically; - check whether Obsidian is installed; - check whether a vault already exists; - check whether Python is available; - conduct the initial short interview; - choose the most appropriate route; - perform the technical setup; - create or adapt the vault structure; - copy the routines; - copy the `vault-lint.py` script; - create the `80_Meta` structure; - perform the final verification. During setup, use the AUTO / PROPOSE / ASK operating levels from the `CLAUDE.md` template as the operating contract. Ask questions directly in chat when you need information from the user. Do not use proprietary, non-portable tools or functions to ask questions. When macro phase A is complete, do not wait for me to request the next step: proactively propose macro phase B, explain what it will do and ask for confirmation before proceeding. ``` Three practical details are worth knowing before you start: - Open Claude Code or Codex in a **clean session** — new, empty and with no unrelated context already loaded. The agent must start from zero to assume the SecondBrain Assistant role correctly. - Choose the **most advanced model available** and enable the **highest reasoning level**. The setup is not trivial work: it involves structural decisions that may shape the next several years of your vault, so you want a model that reasons carefully about each one. - When the prompt begins, the agent clones the bundle into the current folder by itself. You do not need to run `git clone`, download a ZIP archive or prepare anything first. The package is **alive**. We update it when a new variation emerges — as happened with Massimo and will happen again. If you run the prompt again in two months, the agent automatically retrieves the updated repository version; you are not left with an old copy. There is another, subtler point. If the setup reveals a case the bundle does not cover — and it happens — the agent is instructed not to solve it as a local exception inside your vault. Report it to us instead: we will inspect it and decide whether it should become an extension for everyone. Global consistency is what makes the system repeatable. --- ## 10. A final thought about a Makita drill I have used the same metaphor for months because it works: buying a top-of-the-range Makita drill and using it to hammer nails. It works, certainly. But you are using a tool designed to do things that are impossible by hand to perform the same crude task you previously did by hand. AI without SecondBrain underneath is exactly that: a Makita drill used to hammer nails. You perform the same mediocre tasks as before, only faster. AI with SecondBrain underneath does things you would never have attempted manually — reading hundreds of calls to extract a market position, preparing a company profile from six sources in five minutes, or recalling a technical decision from three months earlier when a customer calls without warning. Everything in this article — the four daily cases, the three setup stories and the downloadable bundle — converges on one point: **stop keeping in your head what could live in the system.** Not in order to abdicate responsibility, but to free your mind for the things only you can genuinely do: decide, see connections, speak with people and understand where the company is going over the next 24 months. All the operational material — who told you what, when, why, where a file was placed, and which pattern works for a certain kind of customer — is held by the vault much more effectively than your mind can hold it. Your mind has more important work to do. If you run an SME that feels noisy — incoming contacts that disappear, decisions lost inside WhatsApp, paid tools nobody uses, and customer conversations in which you always feel one beat behind — I understand. We have been there. We now have a way of approaching it that works. Download the package. Open Claude Code or Codex. Write `start`. Give it a few hours of your time, once. Afterwards, the system remains yours for ever. If you become lost during setup or discover something the bundle does not cover, write to me. Genuinely. That feedback is what improves the package for everybody who comes after you. Good work. — *Fabrizio Romano, Futuria Marketing* --- **P.S.** I wrote this article to be useful, not to make it viral. If you reached this point, it has worked at least a little. Now I need your real feedback — not polite feedback, honest feedback. What is clear? What is confusing? Which section did you skip and which did you read twice? Does one case sound too good to be true? What would you like the next article to explore in more detail? Does any part of the setup frighten you or feel beyond reach? Write to me. Reply to the email, leave a comment or send me a LinkedIn message — choose the channel. I use the responses to improve the next articles and the bundle. Not to profile you, place you in an automated sequence or do anything other than understand how to be more useful. Thank you.