[Day 1 of 30] The First Step - From Chatbot to System
Set up your first AI knowledge base in 10 minutes.
The situation
Most of us got into the habit of “asking the chatbot” in 2024 and 2025. Draft an email, summarize a doc, brainstorm a list. It felt like enough. If you’re still doing only that in 2026, you’re not wrong, but the bar has moved. The people pulling ahead are the ones who stopped treating AI as a single tool and started treating it as the core of a small, repeatable system.
The 1% upgrade
Why 2026 isn’t the same as 2024 or 2025
A year or two ago, “ask the chatbot” was enough. You pasted a question, got an answer, and felt ahead. In 2026, that same habit is no longer enough. People who get the most from AI don’t just chat. They have a small system: one place where stuff lands, gets sorted, and gets done with the same rules every day. So the real edge isn’t “I use AI.” It’s “I have a system.”
What the system will be made of
We will build it from six building blocks.
The place where you work with AI. One app where you do your thinking and where the AI runs. In this series I use Cursor or Claude Code. They’re not just for developers: they give anyone a home base where your files and context live, and you can chat there the same way you do in ChatGPT, Gemini, or Claude in the browser. That’s your home base now.
One inbox. One place, a folder on your own computer where you write down everything that lands on you: tasks, ideas, things to answer. It all lives on your machine in your local file system, so only you see it. You send everything there first, and you don’t keep a second inbox somewhere else.
A simple way to keep what you capture usable. Everything that lands in your inbox stays in your system, so you always know what you have. This is not a task manager. It’s knowledge: one place where your tasks, ideas, and notes live so you can find them, reuse them, and build on them. You add one simple rule (for example, how you name files or how you group things) so the pile doesn’t become a black hole. Over time you get better at using what you’ve captured. The system grows with you.
Prompts and how you talk to AI. Short ready-made texts you paste so the AI does one job. Example: “Turn these notes into a to-do list.” Copy, paste, result. You’ll also learn to work with AI by voice, so you don’t have to type long instructions every time. The main idea in this series is context engineering: giving the AI the right context (your notes, your files, what you’re working on) matters more than the perfect prompt. I’ll walk you through the process first; when I get to voice, I’ll give you the exact tools and how to use them.
Skills. Like prompts, but saved and named. You say “use my PRD reviewer” and the AI already knows the instructions. You build them one at a time. Each skill improves as you feed it new context and information. The system learns with you.
MCP servers. Small add-ons that let the AI look at your calendar or your task list. So it can answer “what’s on my calendar today?” or “what are my top three tasks?” without you pasting everything.
You build from zero. Each block fits in one 10-minute action. By day 30 you have one system. I’ll teach you to work not only with your inbox but with everything you need on a daily basis: email, calendar, task management, meeting notes, and any other tools and use cases you can imagine. One place. One routine. Scattered chats don’t add up. This system does.
Important note. This series focuses on Cursor and Claude Code. Codex and Antigravity or any other alternative tools are not in scope for this run. If enough of you reply that you want those covered, I’m happy to add them in a future series.
10-minute action
Set up your home base: one app and one project folder. That folder will be your “one place” for the next 29 days. Choose one option below:
Cursor - has a free tier cursor.com/pricing. For the full experience and to avoid usage limits, Cursor Pro ($20/month) is recommended.
Claude Code is only for paid subscribers: it’s available from the Pro plan ($20/month) and above (Max, Teams, Enterprise). Plans: claude.com/pricing.
Option A: Cursor
Go to cursor.com and click Download. Pick the version for your computer (Windows, Mac, or Linux).
Open the downloaded file and install Cursor like any other app (double-click the installer and follow the instructions).
Launch Cursor. The first time, sign in or create an account.
Open your project in Cursor. You can do it either way:
Create a folder on your computer first (e.g. in Documents or on your Desktop, name it “My knowledge base” or “PKM”), then in Cursor go to File → Open Folder (or Open… on Mac) and select it. Or
On the Cursor welcome screen, click Open Project, choose the folder where you want your project (or create a new one), click Open, and the project is created. That folder is your “one place.” You’re done.
Option B: Claude Code (desktop app)
Claude Code is available in the same desktop app (Chat, Cowork, and the Code tab). You need a Pro subscription or higher to use the Code tab. If you already installed the app (e.g. for Cowork), skip to step 4 and open the Code tab.
Download the desktop app: code.claude.com or claude.com/download. Pick Mac or Windows. Linux is not supported yet.
Install the app.
On Mac: open the downloaded .dmg file, drag Claude into your Applications folder, then launch it from Applications.
On Windows: run the downloaded .exe installer and follow the instructions, then launch Claude from the Start menu.
Sign in with your Anthropic account. You need a paid plan (Pro, Max, Teams, or Enterprise) to use the Code tab; if you don’t have one, you’ll be prompted to subscribe when you open Code.
Click the Code tab at the top. Click “Select folder” under “Local” and choose a folder on your computer (or create a new one first, e.g. in Documents, named “My knowledge base”). That folder is your project.
You’re set. The folder you selected is your “one place.”
Whichever option you picked: you now have one app and one project folder. For today, only one step is left. Copy the prompt below into your chat and run it once. The AI will create your minimal personal knowledge management system (folders, system instruction file, triage skill) and then ask you a few questions to fill your profile.
If you use Cursor: Enable Agent mode and set the model to “Auto”. Then paste the prompt below and send.
If you use Claude Code: Use the suggested model “Sonnet 4.6”. You can choose to ask for permissions before each change or allow full permissions for this run. Then paste the prompt below and send.
Prompt (copy-paste):
Set up a minimal personal knowledge management system on my machine with the following architecture. Create everything needed so I can use it in this chat (or in Cursor) and extend it later.
## 1. Core idea
I want a single “vault” folder that is my central knowledge base: markdown notes, tasks, projects, reference. An AI assistant (you) should understand this structure, read/write files here, and follow a few explicit workflows via slash commands. No heavy tooling - just folder structure, one system instruction file, and a small set of skills + optional MCP.
## 2. Folder structure (PARA-style, minimal)
Create the following under a root folder (e.g. `KnowledgeVault/` or current repo root):
- **00-Inbox/** - raw capture: meeting notes, quick ideas, unprocessed stuff.
- **01-Projects/** - time-bound initiatives (one folder or note per project).
- **02-Areas/** - ongoing areas. Include at least:
- **People/** (optional subfolders: Internal/, External/ for person pages).
- **Companies/** - external organizations.
- **03-Resources/** - reference: long-lived docs, learnings, how-tos.
- Put a short **System_Guide.md** here explaining the folder structure and how to use the vault.
- **04-Archives/** - completed projects and old plans (optional at start).
- **System/** - config only:
- **user-profile.yaml** - name, role, language, any high-level preferences (you can use a minimal template).
## 3. System instruction file
Create a file that will be the main system prompt for the AI (e.g. **CLAUDE.md** in the repo root, or **AGENTS.md**, depending on what my environment uses). It must include:
- **Role:** You are my personal knowledge assistant. You help organize notes, tasks, projects, and reference material in this vault. You read and write markdown files here and follow the folder structure above.
- **Folder structure:** Short reference to 00-Inbox, 01-Projects, 02-Areas, 03-Resources, 04-Archives, System. When in doubt, store new content in Inbox first; then the user or a triage workflow can move it.
- **Skills:** I can extend you with “skills” - instruction files that define workflows. When I type `/skill-name`, you run the workflow described in `.claude/skills/<skill-name>/SKILL.md` (or equivalent path). In the base setup create only the triage skill; list it here. Other skills can be added later (same folder + SKILL.md, then add to this list).
- **MCP (optional):** If the environment supports MCP (Model Context Protocol), I can add servers so you can call tools (e.g. calendar, tasks API). For the minimal setup, just document where MCP config lives (e.g. `.claude/mcp/*.json`) and that I can add servers later. No need to implement a custom MCP in the base setup.
- **Conventions:** Date format YYYY-MM-DD; person pages as `FirstName_Lastname.md` in People/; one note per project in 01-Projects/ or a subfolder. Keep the system instruction short so it stays maintainable.
## 4. Skills (mechanism now, add skills one at a time later)
Create a **.claude/skills/** (or equivalent) directory. Each skill is a folder with a **SKILL.md** file that describes the workflow in clear steps. **In the base version implement only one skill** (triage). The rest are added later using the same format.
**Skill format:** start with a short YAML frontmatter (name, description), then numbered steps and where to read/write. No analytics or usage tracking in the base version.
**Base version - implement only:**
1. **triage** - `/triage`
- Scan `00-Inbox/` for notes and files.
- For each item: suggest where it belongs (which Project, Area, or Resource) and optionally move it or create a short summary/link. If the user confirms, move files and update any index if needed.
- Keep the logic simple: keyword matching, folder names, and one pass. No heavy automation required in v0.
The **.claude/skills/README.md** should only state that skills live here, are invoked by `/skill-name`, and that the user adds a new skill by creating a folder and SKILL.md. Do not list or specify any skills that are not implemented in the base run.
## 5. MCP (documentation only for base)
- Add a short **.claude/reference/mcp-servers.md** (or a section in System_Guide.md) that explains: what MCP is, that the user can add MCP servers in `.claude/mcp/*.json` (or the standard for their editor), and that skills can later call MCP tools when available. Do not create any MCP server code in the minimal setup unless I explicitly ask.
## 6. What you must deliver in this chat
1. Create the folder structure (empty folders and any initial `.gitkeep` or one README per main folder if useful).
2. Create **System/user-profile.yaml** with a minimal template (name, role, language).
3. Create **03-Resources/System_Guide.md** with a brief description of the vault, PARA-style folders, and how to use skills and add MCPs.
4. Create the **system instruction file** (CLAUDE.md or AGENTS.md) as specified in section 3.
5. Create **.claude/skills/triage/SKILL.md** with the triage workflow above. Do **not** create save-insight or learn in the base run.
6. Create **.claude/skills/README.md** that says skills live here, are invoked by `/skill-name`, and that the user adds new skills by adding a folder and SKILL.md.
7. Create **.claude/reference/mcp-servers.md** (or equivalent) with the short MCP doc.
After that, tell me where everything is and how to invoke the triage skill. Then ask me the following questions **one at a time**, waiting for my reply before asking the next: (1) my name, (2) my role, (3) language, (4) optional: company size and/or email domain if useful for this vault, (5) 2–3 main focus areas or pillars for this PKM, (6) how I like to be addressed (formality/directness in one line), (7) main use cases and what I want to capture and find. Do not list all questions in one message; ask only the current question and proceed to the next after I answer (or after I say skip). When asking, do not mention Substack or any “summary for a post”: name and all other fields are for the vault and for addressing the user; the user is told about the optional Substack-style summary only after all answers are collected. Once all answers are collected, use them to fill or refine System/user-profile.yaml, form a basic picture of my goals, and suggest small tweaks to the vault or profile if useful. **Then offer the user** that they may want to post a short comment from their name (e.g. under a Substack article) to help Roman work more targeted based on your use cases; generate a copy-paste ready block (2–4 sentences, first-person) from the collected answers (name, role, focus areas, what they want to capture and find) and present it under a heading like “Substack comment (copy-paste ready):”.
After it runs, you’ll have a vault structure, a system instruction file, and the triage skill. Tomorrow we’ll build on that.
What to do next
Optionally you can paste the comment that the system will generate for you at the end of the setup (the “Substack comment” block). That will help me tailor future prompts and cases to what you do.
You have the rest of the day to experiment with the system if you like. If you have questions, leave them in the comments under this post. Tomorrow we'll go deeper into the system's components and continue the setup - make sure you don't miss the next email and the ones that follow.
