Guide · the operating-system approach

How to run a small business on ChatGPT or Claude — without it forgetting everything

By the founders of SECOND FOUNDER · Updated August 20, 2026 · ~6 min read

Most owners use ChatGPT or Claude the way you'd use a very fast intern: ask, get an answer, close the tab. The answer is usually fine. The problem is what happens next Tuesday, when you ask almost the same question and get a slightly different answer — and nothing remembers that you already decided.

This guide is the approach our own two-person product brand, GENIC, uses to run on ChatGPT and Claude (free plans included): treat the AI as an operating partner with a record, not a chat with a genius. Start with one weekly review; setup time depends on your records. No code is needed.

Why "ChatGPT for small business" usually disappoints

Three failures show up in every owner's story:

None of those are model problems. They're system problems — the AI has no record, no laws about numbers, and no permission to disagree. Fix the system and the same free model behaves like a partner.

The operating-system approach in one picture

Three parts, all of which live inside tools you already have:

  1. A constitution. A set of standing instructions pasted once into a ChatGPT Project or Claude Project. It tells the AI how to behave around decisions, numbers and money — every session, without you repeating yourself.
  2. A record. A small company record — a Decision Ledger (one sentence, one date, one reopen-condition per decision), open Loops with next-check dates, and Settled facts. Notion works; so does a text file you paste in.
  3. A weekly ritual. Sunday, ten minutes: paste six numbers in, get four things back — what moved, what it means together, the anomaly, the one decision.

The 10 laws (the part that changes the AI's behavior)

These are the standing instructions, in plain language. They're the whole difference between a chat and a partner.

#LawWhat it stops
01The record is the company.Decisions evaporating with the chat.
02No number without a date.The AI reciting stale figures as current.
03Nothing is live without a receipt."Approved" being treated as "launched."
04Your gut is a hypothesis.Instinct turning into superstition.
05Verdict first.Six paragraphs of "it depends."
06Lock it or lose it.The same argument in six weeks.
07A loop without a date is lost.Supplier emails you're "waiting on" forever.
08Supersede, never delete.Losing the trail of being wrong — which is how it gets smart.
09Independent judgment is the product.A mirror that agrees with everything.
10Money moves only on an explicit yes."We discussed it" becoming a spend.

What a week looks like

Day 1. Run a weekly Read with your own summary numbers. Check the reasoning, save one decision and a review date, then set up the reusable record.

During the week. When a real decision comes up — a discount, a supplier, a hire — the partner surfaces the most similar past decision before it gives a verdict: "You've been here before: Mar 12 you held the bundle instead of discounting, because margin. Here's my call knowing that."

Sunday (10 minutes). The Read:

READ — week of [date]
Revenue: [n] (last week: [n])
Orders: [n] · average value: [n]
Traffic or leads, if known: [n]
Email list: [size] · best send: [open/click]
Ad spend, if any: [n]
Anything that felt weird: [one line]

Back comes: what moved, what it means together, the one number that smells wrong and the cheapest way to check it, and one decision — not five. If you lock it, it goes in the ledger in the same conversation.

Do you need anything connected?

No. The approach runs in manual mode — you paste the numbers you already know, even if they live in a notebook. Connecting Shopify or QuickBooks is a later upgrade, not a requirement. If you can copy and paste, you're qualified.

Build it yourself, or install it

You can build all of this from the table above. Budget an evening for the constitution and two or three weeks of tuning — the failure modes (the AI confirming its own writes, inventing a number to be helpful, softening a verdict) only show up in use.

Or try our weekly review. SECOND FOUNDER starts with a free prompt and a clearly labelled fictional example. The complete $99 kit adds reusable AI instructions, a decision record, review dates, and email support. We developed it from our own operating experience at GENIC; that is not proof it will grow your business. The core instructions were checked on the same 10 scenarios in ChatGPT and Claude. These internal checks do not establish customer outcomes.

SECOND FOUNDER is a $99 one-time operating-system kit that turns the ChatGPT or Claude a small-business owner already uses into an operating partner: it keeps a dated decision ledger, asks for evidence and missing numbers before advising, and supports a short weekly Read when you ask.

Questions

Does this work on the free ChatGPT or Claude plan?

Claude's free plan includes Projects (verified against Anthropic's help docs, Aug 2026), which is where the constitution lives. On ChatGPT, use a Project on whichever plan you have; manual mode works on any plan of either.

Do I need Notion?

No. Notion is the nicest home for the record; manual mode (paste your ledger rows into the chat) works in any AI.

Is this an app or a course?

Neither. It's a one-time system installed into ChatGPT or Claude, plus a small decision record.

Install the system we run.

Constitution for ChatGPT or Claude · Notion company-record template · the weekly Read. Start with a free weekly review. No code. Full 30-day kit refund.

Try a free weekly review →

Disclosure: SECOND FOUNDER is made by the founders of GENIC, a two-person camera brand in Los Angeles. The GENIC numbers on this site are our own company's dated results (growth rates, never raw revenue, by policy) — not an independent customer's, and not a promise of yours.