Set up an AI that knows your business

“an AI that knows your business is a folder it reads first”

Nate Herk, who runs the AI Automation channel and an AI education community, runs his company from one folder on his computer. His business wiki, his meetings, every video he has made, his skills and his agents all live in it as files, and Codex reads the folder before it answers anything. Ask it for a promo video and it already knows what the community is, what it looks like, and where to grab screenshots. This guide distills how he builds that folder from nothing, the order the parts go in, and the habits that keep it growing.

01Four parts, built in order

Nate sorts everything the folder holds into four parts. The first two make the agent know you. The last two make it do things.

Part	What goes in	How often it changes	Example
Context	Facts about you, the business, your goals	A few times a year	Who buys from you, this year’s targets, how you write
Connections	The tools whose data changes every day	Constantly	Email, chat, project board, revenue, calendar
Capabilities	Skills: written procedures the agent can run	When you learn a better way	How to draft a newsletter in your voice
Cadence	Agents and automations that run on a schedule	When a routine changes	A Monday summary of the week’s meetings

The order matters. A skill or a scheduled agent built before the first two parts exist produces generic work, because the agent has nothing specific to fill it with. Context tells it who you are; connections let it reach for today’s data at the moment a question needs it. Only then does a skill know whose newsletter it is writing.

The split between the first two parts is the one Keep a growing knowledge base accurate calls expertise context and situational context: facts that rarely change stay loaded, daily data is looked up when a question needs it.

02Three tests that it is working

Nate uses three checks to tell whether the folder has reached the point where it changes how he works:

• A teammate asks you a question, and you realize the agent would answer it better and faster, with the sources. The folder holds more than you can remember on the spot.
• You stop opening other tabs. After the setup, force yourself to do everything from the one chat: reply to messages, check the calendar, look up a contract. Each task you do there is one fewer app to switch into.
• Knowledge leaves your head. You no longer rehearse what you decided last quarter or what a customer said in a meeting, because you trust the agent to find it in the chat thread, the email, or the transcript.

03Start with a folder and a map

The whole setup begins with an empty folder on the desktop, opened as a project in the Codex desktop app. Nate’s own starter pack, a set of skills he gives away in his free community, then interviews you with seven questions (your name, what the business does, what you are working on) and writes the answers into a context/ folder as three files: about the business, about you, and priorities. It also creates the one file that matters most.

That file is AGENTS.md, the instruction file Codex reads at the start of every session. Claude Code reads the same kind of file under the name CLAUDE.md, and the two hold identical content, so the pack writes both. Moving between the tools is a copy and a rename.

“Before it even will read your message, it reads this first.”

Nate Herk, “I Turned GPT-6 Astra Into the Ultimate AI Second Brain” · 7:21

Nate’s file is mostly a map. It opens with two lines of identity (you are my operating system, here is your job) and a few rules: be concise, use bullet points, no em dashes, Oxford comma. Everything after that is routing:

• Business advice or team questions: the wiki
• Corporate structure and entities: this section of the wiki
• Nate’s voice and style: the voice file
• Course knowledge: the course folder

His folder holds hundreds of files, and the map is what lets an agent find the right one on the first try. Keep a growing knowledge base accurate shows how to write the map and test it; Turn repeated fixes into rules covers the few behavior rules worth keeping above it.

Connections come next, and the agent can plan them for you. List the apps you open every week (Nate’s checklist: revenue, customers, calendar, messages, tasks, meetings, knowledge), then ask the agent how to connect each one. It will name the API key, the MCP server, or the browser automation that reaches it.

04Interview what is in your head

Once the structure exists, the hard part changes.

“The real issue is getting everything from your head into the system.”

Nate Herk, “I Turned GPT-6 Astra Into the Ultimate AI Second Brain” · 14:04

Automating the small tasks is now easy for the model. What it cannot do is know what you have never written down. Nate’s fix is a skill he calls grill me, adapted from Matt Pocock’s skill of the same name (Pocock’s version interrogates a coding plan before any code is written). Pick a topic, say current priorities, and the skill asks question after question until the topic is exhausted. Every interview writes a new file into the folder, so each session of 20 or 30 minutes leaves a document the agent can read later.

Run it whenever a topic lives only in your head. The interview is faster than writing the document yourself, and the questions surface things you would not have thought to include.

05Let the agent find the links

A pile of interview files and pulled-in transcripts is still a pile. Nate’s wiki shows how each fact relates to the rest of the business, and which entries have gone stale. He gets there with Andrej Karpathy’s LLM wiki, a short public note that describes a pattern. The agent reads your source documents and writes a linked set of pages, one per person, project, or topic. As new sources arrive, it updates the pages instead of starting over.

Using it takes one paste. Start a new chat, say you want a wiki built from everything the agent already knows about your business contacts, and paste the note in. The agent reads the pattern and applies it to your folder.

Nate keeps one wiki per kind of source: one for his videos, one for business knowledge, one for meeting transcripts. A single wiki was enough to start; he split it once each pile began growing on its own schedule. Build a knowledge base your team can ask covers the same move for a whole team.

06Audit, level up, build

Nate keeps the folder improving with a loop of three skills:

1. Audit. Scores the folder out of 100 across the four parts, checks that the two instruction files still match, and writes the report into an audits/ folder. A folder built five minutes ago scored 30.
2. Level up. Reads the latest audit and proposes the next thing to build, tied to the bottleneck you named in the seven setup questions. His demo folder got two candidates: a team action list built from status updates, and a backlog of improvements for his certification program.
3. Build. Set up the connection or automation the level-up proposed. Then audit again.

Because every audit is stored with its date, the scores form a record. Run it every Monday or every 2 weeks and you can see whether the system is improving month over month, rather than feeling that it is. The audit writes a report and waits for you, the same shape Keep a growing knowledge base accurate describes; running it on a schedule is a job for Automate a workflow as a Claude routine.

07The files are the asset

“All you're building here is you're building files and folders.”

Nate Herk, “I Turned GPT-6 Astra Into the Ultimate AI Second Brain” · 18:35

Nothing in the folder belongs to one vendor. Nate built his on Codex, opens the same folder in Claude Code, and can point a new agent at it the day that agent ships. When a better model arrives, the knowledge moves with him; the tool is the part he swaps. Every new repo starts with your lessons makes the same argument for a codebase.

The same logic picks the model for daily use. Nate built his folder with GPT-6 Astra, recorded 4 days after OpenAI announced it, and calls it the most capable model he has used. He also says that for most of what the folder is for (find this document, make this spreadsheet) it is overkill and eats the weekly usage limit faster than GPT-5.6 Sol, which OpenAI prices at $5 per 1 million input tokens. His advice: build with the top model to learn what it can do, then switch everyday questions to GPT-5.6 Sol at a lower effort setting (how hard the model thinks per answer), and turn the top model back on for the hard builds. Recheck your model choices each release covers how to make that call each time the lineup changes.

Further reading

• Nate Herk · I Turned GPT-6 Astra Into the Ultimate AI Second Brain, the video this guide distills
• Andrej Karpathy · LLM wiki, the note to paste into the agent for the wiki step
• Matt Pocock · grill-me skill, the interview skill Nate’s version adapts
• AGENTS.md, the cross-tool instruction-file convention
• Wikipedia · Personal knowledge base, the older idea (a person’s own notes, linked and searchable) that the folder in this guide extends to an agent
• OpenAI · Previewing GPT-5.6 Sol, the model Nate recommends for everyday questions against the folder