What flows into the system
Data is created where the work happens: in meetings, in email, in the calendar. The collects it itself, or receives it in a single pre-approved folder.
Meeting recordings
The recorder captures the meeting; by evening it is text.
The most important input of the whole system. Every evening the agent checks for new recordings, downloads the transcripts and turns them straight into a write-up: decisions, tasks and above all promises (“what I told whom I would send”). The goal: the client has the output within 24 hours.
Email, but not all of it
Only starred threads and mail from people on the .
The agent never reads the whole mailbox. The reading scope is closed: a star in my inbox means “waiting to be dealt with” (until I remove it myself), plus unread mail from approved people. The output is queues, reminders of promises and prepared replies.
Calendar
Reads meetings, watches for clashes and my capacity.
From the calendar the agent can tell what is coming and what has happened: it reminds me of materials before a meeting and expects a transcript after it. It flags when meetings pile up beyond a reasonable limit or creep into the evenings. It may only write blocks marked [NÁVRH] (Czech for “proposal”), and it must never invite anyone.
The “handle it yourself” folder
Whatever I drop here, the agent files without asking.
The only place where “process and file without asking” applies. Articles, notes, materials from clients. Other AI helpers (the one on my phone, for instance) may contribute to the system only through here: one insight, one file.
Downloads
A daily look: what has arrived and where it would belong.
My admitted weakness: I download and never sort. So the agent only takes a look each day (it moves nothing) and presents a filing proposal. The move happens only after my approval: report, consent, move.
Documents and attachments
PDFs and presentations live here; the system links to them.
The knowledge system holds only text. Originals (contracts, presentations, source materials) have one canonical place on Drive and the pages link to them. The agent reads Drive; it may write only to a single pre-approved backup folder.
Who looks after the system
runs on top of the folder. It is not a chat that forgets everything once you close it: its behaviour is set by three files that it loads before any work.
Who you are and what the priority is
The agent’s role, priority no. 1, how every session runs.
The founding document. It defines the role (“steward of knowledge and the wing that follows through for me”), priority no. 1 (outputs from meetings), the procedure for processing new files and how a session starts and ends. When a new rule proves itself, it is added here: the constitution grows from experience.
Values and decision-making
Strengths, vision, weaknesses, “take it / leave it” criteria.
To advise me well, the agent has to know who it is advising. I wrote down my values, a vision for the next two to three years and my weaknesses (“everything tempts me and I say yes”). When I am considering a new project, the agent takes the side of my vision, not of my enthusiasm in the moment, and reminds me what a “yes” would cost.
Guardrails: what it must not do
12 safety rules that stand above any task.
Where it may write, that it never deletes anything, that it only prepares email, that sensitive figures never leave the computer. When a task conflicts with the rules, the agent stops and asks. Trust in the whole system rests right here.
: ready-made procedures
A proven procedure = instructions you can call up at any time.
A meeting write-up, a proposal with a price, a presentation in my visual style, my tone of voice, a text “humaniser”… Each procedure is a file with precise instructions and templates. The rule is: procedures first, creation second. Nothing gets invented twice.
Automatic runs
Morning briefing · evening processing · Friday recap.
Morning: what is ahead of me today, what is pending, what is overdue. Evening: new transcripts into write-ups, backup, updated overviews. Friday: a look back at the week. Every run ends with an entry in the operations log, so you can trace what happened when.
The , a single source of truth
A plain folder on a disk with nothing but inside. No database, no dependence on a single vendor: if Claude and Obsidian both disappeared tomorrow, the texts would stay readable.
Knowledge by topic
Pages on clients, professional know-how, the business.
The heart of the system. Every source is read once and summarised on a page in my own words, with a link to the original. The largest part: clients. Each one has a page with context, history, tasks and promises. Finance pages live under a special regime: only on the computer, outside the online backup.
What happened when
Dated records: write-ups, drafts, reports.
The system’s timeline. Meeting write-ups, drafts of proposals and emails, one-off reports. The rule of layers: the wiki holds living knowledge with no date in the file name, the journal holds everything dated. A trace of the process does not belong among the knowledge.
Operating memory
Operations log, work queue, daily overview, catalogue.
The log is a journal of what the agent has done. The is the queue of what should happen. The dashboard is the home page with tasks. Thanks to them, the next session picks up exactly where the previous one ended.
Processed originals
The source materials the knowledge pages link to.
A long text (a transcript, an article) is read in full only once, when it is processed. After that the work is done with the summary; the agent goes back to the original only on explicit request. It saves time and AI costs.
How the system improves itself
A second brain is not a project that gets finished one day. It is an organism: every mistake and every repetition moves it forward. We built it the way I teach clients to design services: understand first, then a small experiment, and only then scale.
Lessons learned
Every mistake is written down and becomes a rule.
When I return an output with a comment, the agent first records the lesson (what was wrong, what rule follows from it) and only then fixes it. I don’t want to see the same mistake twice. Once a month the lessons are reviewed, and whatever has proved itself is promoted to the constitution or the procedures. In effect formative assessment, just for a system instead of students.
: the end of repetition
Third repetition of manual work → a proposal for a procedure.
The agent watches which tasks I keep doing by hand. When it sees me doing a similar thing for the third time, it suggests: shall we build a procedure for this? That is how the procedure for proposals came about, for example. Bit by bit, routine turns into written know-how.
Ideas don’t get lost
Whatever comes up in conversation goes into the queue with a date and context.
When an idea or task comes up in conversation and isn’t dealt with right away, the agent adds it to the queue itself. At the start of every session it reminds me of deadlines, blocking items and anything that has been sitting for more than two weeks.
Done only with evidence
A task is ticked off only once it is visible that it happened.
The task “send the proposal” is not ticked off on a feeling: the agent checks in the sent mail that the email really went out, and adds the date and a link to the tick. Sent proposals and follow-ups live on as “waiting for a reply”: after a week of silence a draft reminder is created, and after two weeks the item moves to my decisions.
Phases of the build
From rules through a small pilot to automation.
First rules and understanding (an inventory of all the data), then an experiment in one small area, then scaling topic by topic, and finally automation. The same approach as in designing any service: four phases that I teach as Discover, Define, Develop and Deliver.
↻ the loops return to zone 2: lessons change the constitution and the guardrails, repetition gives rise to new procedures
The gate: a human approves. Nothing goes out without me. The agent prepares, I decide: every email, every client write-up, every price. In the AI world this is called ; for me it is simply: a draft is a proposal, not an action.
What comes out of the system
The point of a second brain is not tidy notes. It is speed and reliability towards the people I work with.
Client write-up
A structured write-up of the meeting by the next day.
Priority no. 1. A meeting turns into a write-up with decisions, an action plan and task owners; after my approval it goes out as a document to the client’s folder. The bottleneck is no longer my evening energy, only my approval.
Prepared replies
of follow-ups and replies in the right thread.
The agent creates the draft; I read, edit and send. The Send button is always mine. That way promises from meetings turn into sent emails, not into guilt.
Proposals and documents
Content from the knowledge base, a document in my visual style.
A proposal is built on the client’s page: context, an internal “take it / leave it” mirror, a suggested price (I decide), text in my voice. Final documents live on Drive; the system links to them.
Calendar proposals
Blocks of time to confirm, never invitations on my behalf.
When an email implies “price this by Thursday”, the agent proposes a calendar block marked [NÁVRH]. I can confirm, move or delete it at a glance. The agent must never invite anyone else.
Daily overview
What is on today, what is pending, what others promised me.
The home page: today’s schedule with context, overdue tasks, clients’ promises (those are tracked too), the decision queue. One view instead of five apps.
Articles and newsletter
We write longer texts together, block by block.
The agent is the editor and the memory: it holds the topics, my exact wording and the writing process. The text is created in a joint session, block by block; my sentences must not be rewritten into a “better” version, and I always approve and send the final piece.