THE METHOD
Every Fix ends the same way. You built something. Now you put it in front of an AI and make it earn its place.
You built something: a model, a log, a set of notes. Now you put it in front of an AI and make it earn its place. Not to produce answers. To test what you built, and to find what you have missed.
The Fix is done when the Check passes. This is what you do with a finished Fix, not the last stage of finishing it.
The method is the same every time. What changes is the artefact, and the artefact’s AI Review tab tells you exactly what to hand over, what to ask, which number to check by hand, and where the model will be confidently wrong about your business in particular.
Short of time? Do Prepare, Interrogate and Verify. Ten minutes. Paste it, ask the one question that matters, check the one number by hand. Come back for the rest when the stakes warrant it.
Prepare. Give it the artefact as text, not a screenshot. If you have a Context Pack, load it first. If you do not, the tab gives the minimum context this Fix needs. Say what is actual and what is forecast, what is fact and what is opinion. The sheet knows the difference; the model cannot see it.
Check. Make it read back what you built, in its own words, before it does anything with it. If it has misread a column, everything after is wrong and sounds fine.
Interrogate. Ask for the consequential point: the one number, row or entry that matters most, and why. Not a list. The one.
Stress-test. Change the assumption you least want to be wrong about, and ask what breaks first.
Verify. Two things, every time. The evidence: treat every number the model produces as a claim and re-add the decisive one in the sheet yourself; for documents, find the quotation, the date, the source. If your figure differs from its figure by more than rounding, do not interpret the answer until you have found the cause. The inference: check that the thing it says matters is genuinely the thing that matters, and that the change it points at can actually be made. A model can count perfectly and still recommend attacking work that cannot move. Fluency is not correctness, and the model does not know the difference.
Where it gets this wrong. Every tab carries this block, written for that Fix. It names the specific, observable mistake the model will make about your business with this artefact: the assumption it cannot see, the row it will weigh wrongly, the answer it will call fine. It is the most Fix-specific knowledge in the library, and it is written from running the review, not from theory.
Close. Write one of three things in the tab: what changed, what was confirmed, or what is still open. If something is open, it goes in your Exposure Register. If the result is confirmation, the line in the tab is enough.
The same question, answered twice, on a thirteen-week cashflow.
Fluent: “Your cashflow shows some tight weeks in the middle of the period. Consider chasing receivables earlier and building a buffer against late payment.”
Useful: “Week 7 is the thin week: £3,200 above the floor, and it depends on the £18,000 receipt from [customer] arriving on the 14th. If that slips to week 8 you breach by £14,800. Which of the week 7 receipts is invoiced and which is still a promise?”
The second names a week, an amount and a receipt, and asks you something you can only answer from your business. That is the standard. The first could have been written without your spreadsheet open. Most first answers are the first kind.
Make it specific. Paste the exact rows it is talking about and ask again.
Make it commit. Ask for the number, the date, the consequence. No ranges, no “consider”.
Make it challenge itself. Ask for the strongest case against its own conclusion, and what evidence in your artefact would settle the question.
These are prompts. Good ones. They only work with your artefact in front of you, they demand evidence from it, you are taught to reject a weak answer, and the process ends in a decision you wrote down. That is the whole difference between this and a prompt library, and it is enough.
The method is free and it is the same every time. What it needs is something worth reviewing: a model, a log, a set of notes you built to close a real exposure. That is what the Fix Library is for.