Turn an AI Correction into a Business Procedure

You correct your AI on the same things every week. The discount you never give. The deposit it leaves out. The tone that is not yours.

Each correction works, in that chat. But a correction in one chat is not a saved procedure. Next week, in a new chat or in a different assistant, you are relying on it to come up again.

Here is one way to turn corrections into something your AI can look up. One invented business, one worked example, start to finish.

The business

Sam runs a small garden design studio. Most days, ChatGPT drafts the replies to new enquiries and Claude helps with quotes.

Sam set up the Business Operating Manual template, which gives three places to write things down:

• Start Here: what the studio does, who it serves, the rules that always apply.
• Procedures: how recurring work gets done, one task per note.
• Decisions: prices and policies, with the reasons and the date.

Rules that hold for every customer go here. What one customer asked for belongs in a client note instead, the way freelancers keep client context between ChatGPT and Claude.

During setup, ChatGPT helped Sam write one real procedure: "Reply to a new enquiry". Four steps. Thank them, check they are a fit, use the current price from Decisions, end with one next step.

Week one: the reply that forgot the deposit

A customer asks whether Sam can design a small courtyard this autumn. Sam pastes the email into ChatGPT:

Check my operating manual, then draft a reply to this enquiry.

The draft is good. It is friendly, it confirms the autumn slot, it quotes the design fee from the pricing decision. It does not mention the deposit, because nothing in the procedure says to.

Sam tells it: we always take a 30% deposit, mention it.

Making it stick

This is the moment that usually gets lost. The folder instructions ask the AI to tell a standing change apart from a one off. So ChatGPT asks: is that how you always want it, or just for this reply?

Sam says always. A number belongs in Decisions, because the procedure tells the AI never to quote one that is not there. So ChatGPT saves "Decision: 30% deposit to confirm a slot", then adds a rule to the procedure that points at it, and says what it changed:

## Rules
- When confirming a slot, mention the deposit from the current deposit decision.

Sam opens both notes and checks the edits. If the AI had rewritten more than it should, note history shows exactly what changed and can restore the earlier version.

Had Sam said "just this once", the procedure would stay as it was. A manual that rewrites itself on every passing remark stops being a manual.

Week two: a different chat

Another enquiry. New chat, and this time Sam uses Claude:

Check my operating manual, then draft a reply to this enquiry.

The folder instructions ask the AI to name the procedure it used, and Claude says it followed "Reply to a new enquiry". The deposit is in the reply.

That is the payoff, but check it rather than trust it. ChatGPT and Claude decide when to call a connected tool, so ask which procedure it used, and look at the tool activity to confirm it actually read the note. If it did not, ask it to.

Month three: the price goes up

Sam raises the design fee for new customers from January.

The old pricing decision is not quietly overwritten. The Decisions folder instructions ask for a new note when a decision replaces an old one, with the two linked and the old one tagged "superseded":

# Decision: design fee from January

## Decision
From 1 January 2027, the design fee for new enquiries is €1,200. Until then it is €1,000. Quotes already sent keep the amount quoted.

## Why
The old fee no longer covered site visits.

## What we gave up
Some smaller courtyard jobs will go elsewhere.

## Date
October 2026

The reply procedure never contained a price. It says "use the current pricing decision", so nothing in it has to change. The instructions tell the AI not to apply a superseded decision, to apply the new one's dates and exceptions, and to ask Sam if the rate is missing or two current decisions seem to conflict.

That split is the useful part. Procedures hold the how. Decisions hold the numbers and the why. It is the decision log recipe with one addition: the decisions are also the numbers your AI is told to use.

What the instructions can and cannot do

Everything above is what the folder instructions ask for. They are instructions, not guarantees. They also tell the AI not to invent prices or policies, and to ask when the manual does not say.

What makes it work in practice is small: ask for the manual by name, ask which procedure it used, and read the edit when it updates one.

Try it with your last correction

1. Connect Hjarni to ChatGPT or Claude.
2. Paste the link to the Business Operating Manual template into a chat and ask your AI to set it up. It creates the folders and three starter notes, and the instructions tell it to ignore the starter content until you replace or confirm it.
3. Tell it the last thing you corrected, and ask it to turn that into your first procedure.

Pick the task you keep correcting. Set up the Business Operating Manual, then make your last correction its first procedure. Free to start.