Your values, decision criteria, and ways of thinking — kept on your own computer, in plain files, growing every week. AI works from your context instead of handing you the same generic answer it gives everyone else.
You think you are deciding for yourself. Often you are following a recommendation: the feed showed it, the assistant suggested it, you picked it. Repeat that a few hundred times and the line between your judgment and the AI’s suggestion disappears.
Owning your own data is where everything starts. Not in someone’s app. Not in a chat history you can’t search. In files you can open, read, back up, and move.
Paste these into your AI agent one at a time, in order. Replace everything in 【brackets】 with your own words before sending. Read every file the AI produces; if anything is vague or doesn’t sound like you, tell it to fix that before moving on. The reasoning behind each step is in Steps 1–6 below.
① Create the workspace (Step 1)
This folder will be the workspace where I run my business with AI.
Create the following folders. Do not create any files yet.
inbox/
ontology/concepts/
ontology/people/
ontology/organizations/
ontology/relations/
actionlog/
When done, show me the folder list.
② Put a rules file in place (Step 1)
Create AGENTS.md in this folder.
(If the AI you use reads CLAUDE.md or GEMINI.md instead, put the same content under that name too.)
Rules to write:
- Work only inside this folder
- Ask me before deleting files, sending anything outside, or publishing anything
- Files in ontology/ start with a header block (id, type, sensitivity, owner, updated)
- Never put test data or jokes into ontology/. Experiments go in inbox/
- If something can't be verified, write "unverified." Never fill gaps with guesses
Keep it short, as bullet points. Show me the full text when done.
③ Get your thinking out (Step 2)
I'm going to talk (or paste) freely about 【topic: e.g. my strengths / which business to focus on】.
Save it to inbox/ as a file with today's date.
Don't over-polish it. Keep my own wording as much as possible.
【paste what you said — a brain dump, a voice memo transcript, a photo of handwritten notes】
④ Turn it into a concept file (Step 3)
Read 【file name】 in inbox/ and draft a concept file in ontology/concepts/.
Topic: 【e.g. how I choose my market】
Include:
- What I currently think (in my words)
- People and organizations involved (create files in ontology/people/ and ontology/organizations/ if needed)
- Open questions I haven't decided yet
- The header block (id, type: Concept, sensitivity, owner, updated)
Do not add general advice. Do not write anything I didn't say.
Then read it and send it back with corrections — this back-and-forth is S1:
I've read the file. Please fix the following:
- 【This phrase isn't how I talk → use plain everyday wording】
- 【I don't understand this part → add a concrete example】
- 【Add this: ...】
⑤ Log today (Step 4)
Add today's actions to actionlog/【YYYY-MM】.md.
Check today's date from the system clock.
Format: | Time | Type | Business | What I did | object_id | Result / note |
Today:
【e.g. 09:00-10:00 Thinking · Bakery · Reviewed pricing → decided to raise prices 10% next month】
⑥ Map the connections (Step 5 · weekly)
Read the whole of myproject and map the connections between people, organizations, concepts, and the action log.
- Concepts that relate to the same theme
- Concepts tied to a specific person or organization
- Recurring interests or decision patterns visible in the action log
Propose additions to ontology/relations/links.json. Show me the proposal before writing anything.
⑦ Analyze the whole map (Step 6 · monthly)
Read all of myproject and analyze it from the S1 point of view:
- Themes I keep coming back to
- Gaps between what my concept files say and what my action log shows
- Decision criteria I use but haven't put into words
Most AI apps are black boxes. You can’t see what they keep or how they decide. S1 aims for the opposite: you always know where your data is and what the AI is reading. Everything lives in plain files on your own machine. If a service shuts down tomorrow, your data is still yours.
You need one AI agent app that can read and write files. An AI agent is an AI that works inside a folder you point it at — it reads files, creates them, and rewrites them. Unlike a chat window, the result stays on your disk as files.
Choose one that…
Why
lets you choose the working folder
so it sees myproject and nothing else
can ask before it edits or deletes
so you can approve each change at first
runs on a subscription you already have
so you don’t add another bill
You never need to write Markdown by hand — ask the agent to “make a concept file on this topic.” A text editor to read the files yourself is handy but optional.
As of October 2026: popular choices include Claude Code (Anthropic, with a desktop app), Codex (OpenAI, included with ChatGPT plans), and Antigravity (Google, an editor with Gemini built in). The steps in this framework are the same whichever you pick. Names, prices, and features change every few months — choose by the criteria above, not the brand.
Tip: start a fresh conversation for each piece of work. One conversation stretched over days gets slow and can stall. Because everything important is in files, closing a conversation loses nothing. That is one more reason to keep your data in files, not in chats.
Create the folders. For S1 you need only three: inbox, ontology, actionlog. Name the top folder whatever you like — this handbook calls it myproject.
myproject/
├── AGENTS.md rules the AI reads every time it starts work
├── inbox/ voice memos and rough notes land here first
├── ontology/ the map of your head
│ ├── concepts/ 1 theme = 1 file
│ ├── people/ 1 person = 1 file
│ ├── organizations/ 1 organization = 1 file
│ └── relations/ how things connect (links.json)
└── actionlog/ what you actually did, month by month
Tip: short, lowercase English folder names. They make instructions shorter and avoid tool quirks. inbox = “throw it here first,” ontology = “my map,” actionlog = “what I did.”
As you progress, more folders appear — research/ in S2, content/ in S3, ontology/finance/ in S5, customers/, projects/ and outsource/ in S6. You don’t create them now.
Put a rules file in place first. AI agents automatically read a Markdown file with a specific name in the working folder every time they start work. Whatever you write there is followed without you repeating it.
File
Read by
AGENTS.md
the common name most agents read — start here
CLAUDE.md
Claude Code
GEMINI.md
Gemini (Antigravity)
Five lines are enough for S1 (prompt ② above): where the AI may work · what needs your approval · how files are written · real vs. test data · never guess. Grow it one line at a time whenever you notice “I keep saying this.” In S3 you add automatic logging.
Give every file a header block. Files in ontology/ start with a small header (frontmatter):
---
id: concept-market-choice
type: Concept
sensitivity: internal
owner: me
updated: 2026-10-11
---
With sensitivity (who may see it: public / internal / restricted) and updated, you can later ask “show me files not updated in six months” or “list everything restricted.” In S6, when you start holding data about other people, the same header does the heavy lifting.
Keep a change history. The AI rewrites files. Work somewhere you can roll back — your cloud drive’s version history, or Git (an app such as GitHub Desktop makes it easy). If “put it back the way it was yesterday” is possible, you can let the AI move boldly.
① Show the AI only the working folder. Never point an agent at your whole drive. Move anything it needs into myproject. “The AI only sees the workspace” prevents accidental edits and deletions elsewhere.
Tip: one folder per project. If several pieces of work share a folder — briefs, images, HTML, revision notes — the AI starts reading the wrong project’s files. For a web page or a deck, make projects/cafe-landing-page/ with plan.md, index.html, and img/ inside.
② In S1, you start every run by hand. No “wake up at 7 a.m. and log things automatically” yet. Scheduled jobs burn resources and can spray files around when something goes wrong. Give the instruction yourself each time. Later, once your system is mature, scheduled runs begin with read-only work — a morning summary, a dashboard refresh (S6). Writing on a schedule comes after that.
Every day or every week, put your thinking into inbox/.
Record a voice memo on your phone — just talk about what you noticed or decided.
Turn it into text (most AI assistants transcribe audio files directly).
Save the text into inbox/ — through your cloud drive, or by pasting it to the agent (prompt ③).
Keep each recording to about 10 minutes; very long recordings blur the structure.
Tip: Markdown, CSV, or JSON. PDFs and spreadsheets are expensive for agents to read and lose structure. Convert to text before filing.
Tip: never put test data into the real map. Jokes, fake names, scratch text go in inbox/ or a temporary folder. ontology/ is what the AI treats as fact.
The brain-dump sheet. Paper works too. Fold a sheet of letter or A4 paper lengthwise to make a long, narrow column. Pick one topic — “my strengths,” “what to focus on this year,” “why this market” — and write bullet points top to bottom without stopping and without editing yourself:
Topic: why I want to grow the tutoring business
- Kids can use AI now, but many just copy answers
- I want them to judge for themselves
- Parents are scared of AI
- So safety and good habits have to come with it
- Grades matter too
- But what really matters is owning how you learn
It is not meant to be good writing. It is meant to get what’s in you out, uncensored. Photograph it or read it aloud into a voice memo, and file it in inbox/.
Step 3: turn it into concept files (the logic layer)
From the inbox, build concept files in ontology/concepts/ — one theme per file: my-strengths.md, how-i-choose-markets.md, decision-criteria.md, things-i-avoid.md.
A concept file holds what you think about the theme now, your motivation and confidence, and related events and decisions.
Let the agent draft; you do the thinking. Concept files need headers, tags, links to people and organizations, and history. Typing all that by hand kills the habit. Tell the agent the core — what the file is about, what you think, who’s involved — and let it structure (prompt ④).
Then read every word. Anything that isn’t your phrasing, anything you don’t understand, anything that feels off — send it back. A file you haven’t read, don’t understand, or that isn’t in your words is noise in your Ontology. Don’t keep it.
Tip: count files after bulk work. Hand an agent a whole folder and it may generate far more files than you asked for. Check the count and delete anything you didn’t request — immediately, not “later.”
AI on its own produces documents that look finished and do nothing. What gives your Ontology value is what you actually think, doubt, and decide right now. Use the agent as your editor and structurer; keep feeding it your own thinking.
Step 4: log what you actually do (the action layer)
Record in actionlog/ what you did each day — not just “did I do what the AI suggested,” but where your time actually went: thinking, documents, meetings, research, email, admin, travel, publishing.
One file per month (actionlog/2026-10.md), one block per day:
## 2026-10-12 (Mon)
> Hours: 6 | Main businesses: Bakery, Catering
| Time | Type | Business | What I did | object_id | Result / note |
the id of the related person / organization / concept, or —
Result / note
what was decided, made, or learned
The point is not a list of chores. It is that, later, the AI can see which actions connect to which people, organizations, concepts, and decisions. Log what you did, what you didn’t, and what you did instead — the reasons you didn’t are data too. Over time the AI can see your patterns: where your hours go, what you put off, which actions actually produce results.
Tip: tell the AI when a project stops. Log “paused” or “cancelled,” or update the concept file’s status. Otherwise later analyses keep treating dead projects as alive.
Don’t hand-draw the relationships. Ask the agent (prompt ⑥). As files accumulate, links appear that you never noticed — “my strengths” shaping “how I choose markets,” “things I avoid” pointing at “what to outsource.”
Do this weekly or monthly, and review every proposed link. Remove links that don’t ring true. The map isn’t something the AI makes for you; it’s something you grow together.
Once you have inputs, concept files, a log, and links, ask the agent to analyze the whole thing (prompt ⑦). This is not “asking the AI for advice.” It is rebuilding your own decision criteria from the record you’ve kept.
Read the analysis. Push back where it’s wrong; ask why where it’s surprising. The aim is not for AI to decide for you — it is to update your own judgment using your own material.
To finish S1, build one thing with your agent — a web page, a company description, a slide deck, a small app.
The habit to learn: before building anything, write a detailed brief in Markdown. “Make me a website” produces something plausible and not what you pictured. The brief is where everything in your head goes.
1. Draft the brief.
I want to make 【e.g. an intro page for my café】. First, write a brief as a Markdown file.
Requirements:
- Who it's for: 【e.g. people aged 30–50 who live nearby and haven't been in yet】
- What it should say: 【e.g. we make everything from scratch each morning】
- How it should feel: 【e.g. honest, clean, unpretentious】
- Must include: 【e.g. hours, address, how to book, three signature items】
Sections of the brief: purpose, audience, page structure, content of each section, tone, open questions.
Tip: say which image goes where, and why. Putting photos in img/ isn’t enough. Write in the brief “this photo, in this section, to give this impression.” The AI won’t infer the purpose of an image.
2. Read every line and fix what doesn’t match your picture. “That doesn’t sound like us.” “The thing I most want to say ends up at the bottom.” Kill every mismatch now — fixing it after the build costs several times more.
3. Build from the brief.
Read this brief and build exactly what it says, as a single HTML file.
Do not add anything that isn't in the brief.
Tip: forbid extras. Left alone, AI adds taglines, FAQs, invented testimonials, extra sections. If you want something new, add it to the brief first, then rebuild.
4. Check the result, fix it, and write the fix back into the brief — the important part.
## Notes after building (2026-10-12)
- Photos were dark; "clean" didn't come across → swapped for daylight shots
- Item names alone didn't show the effort → added prep time to each item
- Lesson: put photo direction in the brief, not just words about "feel"
Now the brief isn’t a throwaway spec. It’s a file where your production judgment accumulates — the same record → analyze → update → write-back loop as the Ontology. You’ll use this pattern at full scale for landing pages in S3.
Tip: publish your HTML for free. The easiest route is Cloudflare Pages direct upload: create a free account, upload the folder with your HTML from the dashboard, and you get a https://your-name.pages.dev link you can open on your phone within minutes. Send it to a friend or show it in a sales meeting — seeing your own work live on the web changes how you think about what’s possible.
I want to publish my HTML file for free with Cloudflare Pages direct upload.
Walk me through it one step at a time, including exactly where to click.
I'll send you screenshots of my screen as I go.
Anyone with the link can see the page. Never publish customer data, internal numbers, or personal information. Auto-publishing from GitHub comes in S3.
Done when:
You wrote a brief before building, read it fully, and fixed it to match your picture
You built the output from the brief
You wrote what you changed, and why, back into the brief
(Stretch) You published the page and sent the link to someone
Using an AI agent app, create myproject, the rules file (AGENTS.md), and a few files with headers: a self-introduction, one concept, one project overview
2
Action log and the map
Set up actionlog/ and start logging. Have the agent read the whole folder and organize concepts, people, and relationships
3
Brief-first building
Write a brief, check and fix it, then have the agent build a web page or document from it — and write the fixes back
After three sessions, participants have their own AI-readable workspace, can build from a brief, and know how to feed corrections back into the record.
SOVREN Framework is open-source. The English edition covers S1–S6; S7–S9 and the milestones are published in Japanese. The method is practiced in weekly live classes in Yokohama, Japan, with a report after every session.
Where it is practiced
SOVREN Framework is a free, open canonical text. It is tested in live classes in Yokohama, Japan, where people implement it in their own businesses with an instructor.