Field notes
How I run my company on WhyVault.
I build several ventures side by side, mostly alone, always with AI. Here’s how my vault is set up, how decisions get into it, and why my agents rarely have to ask what we’re actually doing.
By Kilian, founder of WhyVaultReading time: 9 minutesUpdated:
01
Why a founder needs an operating system AI can read
When you start a company alone or with one co-founder, the company lives in your head. Why the price is what it is. What you learned on the last customer call. Which side project is quietly dead. That works fine – until you’re building three things at once.
Then AI joins in. Claude helps me think, Claude Code writes code, ChatGPT handles the rest. Each one is smart, and each one starts from zero. For months I pasted the same paragraphs into chats: who I am, what WhyVault is, what we already decided.
So to me, a company operating system is nothing grand. It’s a folder of Markdown that describes the company well enough that a new person – or an agent – can join the conversation after five minutes. A second brain for founders, open to every agent I allow in. Some people call that an AI chief of staff. I call it good notes.
02
The one rule: the vault answers why
Everything else follows from this sentence. The vault answers why. The task tool answers what’s next.
| – | Vault | Task tool (for me: GitHub Issues) |
|---|---|---|
| Answers | Why are we doing it this way? | What’s next, and who’s doing it? |
| Example | “Agents can’t delete over MCP, because trust is the product.” | “Finish the trash view in the app by Friday.” |
| Lifespan | Years | Until it’s done |
| When it’s finished | Stays, so you understand it later | Gets closed and forgotten |
| Who writes | Me and my agents | Me, my agents, maybe a team one day |
- Answers
- VaultWhy are we doing it this way?
- Task tool (for me: GitHub Issues)What’s next, and who’s doing it?
- Example
- Vault“Agents can’t delete over MCP, because trust is the product.”
- Task tool (for me: GitHub Issues)“Finish the trash view in the app by Friday.”
- Lifespan
- VaultYears
- Task tool (for me: GitHub Issues)Until it’s done
- When it’s finished
- VaultStays, so you understand it later
- Task tool (for me: GitHub Issues)Gets closed and forgotten
- Who writes
- VaultMe and my agents
- Task tool (for me: GitHub Issues)Me, my agents, maybe a team one day
My test for every note: does it have an owner and a date? Then it’s a task. Will someone still need to understand it in six months? Then it goes in the vault. Put both in one tool and you end up maintaining two truths – and the worse one wins.
03
What my vault looks like
This is the WhyVault template, the way I use it myself. Every venture is a room, and every room starts with exactly one file.
0_Inbox/
1_Knowledge/
People/ Software/ Principles/ …
2_Calendar/
Reviews/
3_Rooms/
Ventures/
WhyVault/
00 Overview – WhyVault.md
01 Decisions/
03 Knowledge/
05 Posts/
<Side project>/
00 Overview – <Side project>.md
_Closed/
8_Attachments/
9_System/
Identity/
USER.md SOUL.md IDENTITY.md CONTEXTS.md
Templates/3_Rooms/Ventures/- One folder per venture. WhyVault has grown and has sub-folders; a fresh side project has nothing but its overview. Folders appear only when I need them.
01 Decisions/- Every decision that carries weight gets its own note, from a template. This is my decision log.
05 Posts/- Drafts for LinkedIn and Reddit that an agent writes from the notes in the room. I edit and post them myself.
9_System/Identity/USER.mdsays who I am,SOUL.mdhow I sound,IDENTITY.mdwhat role the assistant plays.CONTEXTS.mddescribes each world under3_Rooms: what it is, where its tasks live, how I measure success, who decides. I didn’t write them by hand – Claude interviewed me and drafted them._Closed/- The only status a room has is where it sits. Done means moved, not deleted.
# Overview – WhyVault## PurposeA hosted Markdown vault that any AI agent can read and write. First for me, then for everyone who works with several agents.## Boundaries- No teams, no built-in AI chat.- Tasks live in GitHub Issues, not here.## Who decides what- Product, pricing, positioning: me.- Agents suggest and draft. They don’t decide.## Rules of the game| Milestone | Continue if | Stop if || --- | --- | --- || After 6 weekends | I use it daily myself | I don’t use it myself || Landing page, 4 weeks | ≥ 150 on the waitlist | < 50 on the waitlist |## Timeline- First version, just for me- [[No deleting over MCP]]## Open questions- Which templates do people need besides PARA?
04
The overview: one page every agent reads first
Roughly what the WhyVault overview looks like. Short, honest, no tasks.
Purpose and boundaries come first. An agent that knows what the project is not suggests less nonsense.
Who decides what sounds silly in a one-person company. But it’s exactly the line that stops an agent from simply making decisions.
The rules of the game, kill criteria included, live here – not in my head. So every agent knows them too.
The timeline links to decisions. It’s the room’s table of contents.
05
How a decision gets into the vault
Most decisions happen in the middle of work. For me, often in Claude Code, inside the repo, between two commits:
You: We’re sticking with it: agents can’t delete anything over MCP. Write that up as a decision in the WhyVault room – including what we assumed.
- Tool: vault_overview
Claude Code: I’ve read the overview and the decision template. The room has no decisions yet, so I’ll create the first one in `01 Decisions/`.
- Tool: create_note · “No deleting over MCP”
Claude Code: Done – three options, the reasoning and the assumptions. The key assumption: a wrong delete often goes unnoticed for days.
You: Add it to the timeline. And the actual work?
- Tool: update_section · “Timeline”
Claude Code: The timeline has a new line with a link. I opened a GitHub issue for the implementation – that doesn’t belong in the vault.
# No deleting over MCP## QuestionShould agents be able to delete notes over MCP?## Options| Option | For | Against || --- | --- | --- || Allow deleting, with a trash | Agents can clean up | Mistakes surface late || Delete after confirmation | I stay in the loop | People click through prompts || No deleting over MCP | Nothing gets lost | Cleaning up stays manual |## DecisionNo deleting over MCP. Deleting only happens in the app.## ReasoningIf you let an agent write, you need to be sure nothing disappears. Trust is the product.## What we assumed- Agents sometimes mix up notes with similar titles.- Cleaning up is rare and can wait.- People care more about safety than convenience.## Consequences- There is no MCP tool for deleting.- I move finished rooms myself.
06
My decision log template
What comes out is a note I’ll still understand a year from now. More in the decision log guide.
Options as a table force me to write down the paths I’m not taking. Those are the first thing I’d forget.
What we assumed is the most important section. When an assumption breaks, I know which decision to revisit.
Consequences turn an opinion into a rule – one the agents follow too.
07
My agents and their jobs
No agent knows more than the others. They all read the same vault – that’s the whole trick. The setup guides show you how to connect them.
Claude on the web & phone
My thinking partner. This is where I sort ideas, prepare conversations and have Claude draft LinkedIn posts in my voice from the note “Positioning & message”. They land in
05 Posts/.Claude Code in the repo
Writes code – and writes back why. Decisions and positioning go into the room. Claude Code condensed the repo’s
VISION.mdinto that “Positioning & message” note, for example. SetupChatGPT
For everything else: a second opinion, quick drafts, another angle on the same notes. Same knowledge, different tool.
Permissions per connection
You can revoke any connection at any time, and set access tokens to read-only. Every change shows up in the history with the agent’s name and can be undone. No agent can delete. More
08
The 15-minute weekly review
Fridays, with coffee. The result goes into 2_Calendar/Reviews/. When I’m short on time, Claude reads the timelines of the active rooms first and drafts it for me.
- Empty
0_Inbox: every note gets a room, moves to1_Knowledgeor goes. - Extend the timelines of active rooms: what happened this week, which decision was made?
- Move finished rooms to
_Closed/. - What really happened? Not what I planned.
- What did I learn?
- What will I do differently next week?
- Where did I get stuck? Usually the most honest line.
09
Side projects with an expiry date
Founders rarely fail for lack of ideas. More often it’s too many ideas that never end. So every side project of mine gets a time limit and a kill criterion before it starts. For example: “After six weekends: am I using it every day myself?” Or: “Landing page, four weeks: at least 150 waitlist sign-ups. Under 50, I stop.”
What matters isn’t the number, it’s the timing. Before, I’m honest. After, I’m in love. Because the rules sit in the room’s overview, every agent knows them too – and can remind me when I’m still polishing the landing page in week eight.
And if a project carries on, just not with me? That’s where the boundary rule helps: rooms only link to 1_Knowledge, never to other rooms. I write a 99 Handover – <Room>.md and pass on the folder. It’s complete, because it never depended on anything else.
My task list tells me what to do. My vault tells me why.
?
Frequently asked questions
Why are there no tasks in the vault?
Because tasks go stale fast and reasons don’t. A vault full of finished to-dos is hard to read, for you and for every agent. Tasks belong in a tool that can close them – for me that’s GitHub Issues.
Is this worth it with just one project?
Yes. You start with one room and one overview. It pays off with the first decision you don’t have to explain again three months later.
Which AI agents can I connect?
Any that speak MCP: Claude on the web, desktop and phone, Claude Code, ChatGPT, Cursor and many more. They all use the same address and read the same vault. All setup guides
What if an agent writes nonsense?
Every change is a version with the agent’s name on it, and you undo it with one click. Deleting over MCP isn’t possible at all, and you can revoke any connection at any time. History and permissions
Can my team work in it too?
No, WhyVault is built for one person today. What you can do: hand over a room as a folder, or export the whole vault as a zip. The export opens directly as an Obsidian vault.
What does it cost?
WhyVault is free up to 20 notes, no credit card needed. Pro is €9.99 a month or €99 a year, with no limit. Pricing
Keep reading
All guidesDecision log
A template for decisions you’ll still understand six months from now – with a filled-in example to copy.
ReadLLM Wiki
Karpathy’s idea explained: an AI maintains your knowledge as a Markdown wiki – and how to get one without scripts.
ReadObsidian & Claude
How to connect your Obsidian vault to Claude and ChatGPT, locally or in the cloud.
Read
Start with one room.
Create your vault from the WhyVault template, write your first overview and connect Claude. Free, no credit card.