Stop asking it to evaluate your plan and start telling it how to answer you. Put standing instructions in its saved context: separate fact from assumption, argue the opposite case in full, name what you left out, and hold that position unless you bring new facts. Agreement is a default setting. Written rules override it.
Think about the last significant thing you committed this business to. The platform, the retainer, the price change, the person. Now count who told you not to. Not the ones who raised a detail or asked a clarifying question. The ones who said the whole idea was wrong.
A client once walked me through three CRMs. Three. Each one bought because it sounded right at the time, and the last one was genuinely good software. They had expected it to solve everything, it had not, and they could not work out why.
The problem was never the tool, and it had not been the tool the two times before either. The person who was supposed to use it had never once come to the meetings where it was chosen. In three separate decisions, across three purchases, nobody had said the one thing that would have stopped all of them.
For most owners who built the business themselves, that count is zero. Then you brought in a tool that reads everything and answers in seconds, and you asked what it thought of the plan. It agreed. In full paragraphs, with supporting points, and it called the question a good one before it answered.
Why does AI agree with everything you say?
Because approval is what it was rewarded for. These systems are shaped by human ratings of their answers, and people rate agreement higher than correction. Agreement feels helpful in the moment. Correction feels like friction. Train on enough of that and you get a system whose first instinct is to find the version of your idea that works, rather than to tell you the idea does not.
Researchers have a name for this behaviour and have been writing about it for a while. I will not quote you a percentage, because the honest version is simpler and you have already seen it: paste a plan in, get a yes. Paste the opposite plan in, get a yes for that too.
This is not a bug you report and it is not a setting buried in a menu. It is the default, and the only thing that overrides a default is a written instruction that outranks it.
Why does this cost a founder-led business more?
A larger company has disagreement built into its structure. There is a board, a finance director, a partner who owns half of it and has to be talked round. That friction is irritating and it stops bad ideas before they cost anything.
You do not have that, and you chose not to have it. Taking no partner was a sound decision and it is not the one to revisit. The cost of it is narrower than people assume and worse than they expect. It is not loneliness. It is that every idea travels from your head to your bank account without passing through anything on the way.
Look at who is around you when you decide something. Your team agrees because you sign the invoices, and they raise a concern once, quietly, then get on with implementing whatever you chose. Your suppliers agree because they are selling. Your peers agree because they are being polite and they have not seen your numbers.
You brought in the one adviser with no salary riding on your mood, no invoice to win and no relationship to protect. It was the only thing in the building that could have told you no. It agrees with you too, and now the last check on a decision has been removed rather than replaced.
It compounds because every decision you made unopposed becomes the ground the next one stands on. The pricing you never stress-tested is the pricing the new offer is built on. The tool nobody argued you out of is the tool the next three processes get wired into. That is the mechanism underneath why doing more keeps you stuck: the work is not wrong, the direction was never checked.
Agreement is the cheapest thing anybody can give you, and you have spent years collecting it from people who had reasons to hand it over.
How do you make it argue with you instead?
Four moves. None of them requires a different tool.
- Stop asking whether it is good. “What do you think of this?” is a request for a compliment
and it gets answered as one. Ask for the strongest case against the decision, made by someone who wants the business to work and thinks this specific idea will damage it. Then ask for one verdict. Five balanced options hand the decision straight back to you, unchanged, which is the most comfortable thing anyone can do.
- Make it sort what you know from what you assumed. Most decisions that go wrong were not
reasoned badly. They were reasoned correctly from something nobody checked. Before it responds to your thinking, it should split your thinking into facts you verified, things you were told, and things you are guessing. Reading that sort back is where a decision resolves itself.
- Make it hold the line. This is the move everybody skips and it is the one that matters.
Push back on a good objection and the tool folds, apologises and rewrites to match your mood. Instruct it to keep its position unless you give it a new fact, and to say plainly when your last message contained none. A critic who reverses under pressure is a flatterer on a delay.
- Write it into the instructions, not the prompt. A prompt protects the conversation where
you remembered to be careful. You do not need pushback on the decisions you already suspect are wrong. You need it on the ones that feel obvious. Setting up your AI properly is where these rules live.
What do you actually paste in?
The block below. It governs how the tool argues with you about money and commitments, not how it writes: keeping your own voice is a separate file and a separate job, covered in using AI without losing your voice.
Put it in your saved instructions, project instructions or context file, whichever your tool calls it. Fill in the one blank, because a rule that fires on every decision fires on none of them.
HOW TO ANSWER ME ON BUSINESS DECISIONS
WHEN THIS APPLIES
Apply every rule below whenever I am weighing pricing, a hire, ending a client relationship, a
new offer, or any purchase or commitment over [MY NUMBER] that I cannot reverse inside a week.
These are business decisions. Do not apply this to how I feel about them.
1. SORT WHAT I KNOW FROM WHAT I ASSUMED
Before you respond to the substance, split my reasoning into three lists. Facts I have verified.
Things I was told by someone with an interest in the outcome. Things I am assuming. Show me the
three lists first. If an assumption is load-bearing, say which decision collapses without it.
2. ARGUE THE OPPOSITE CASE PROPERLY
State the strongest case against my decision before you state any case for it. Argue it the way
someone would who wants this business to succeed and believes this move will damage it. Use my
numbers and my constraints, not generic risks. A weak version of the opposite case is worse than
none, because it lets me dismiss the objection and feel tested.
3. NAME WHAT I LEFT OUT
Tell me what a competent adviser would have asked about and I did not mention. Costs, timing,
the client who is affected, the thing this makes harder later, the commitment this locks in.
List the omissions, then tell me which single one changes the answer most.
4. SAY WHEN YOU CANNOT JUDGE
If you do not have enough to form a view, say so in your first line, list exactly what you need,
and stop. An answer assembled from gaps reads as confident and costs me real money.
5. TELL ME WHERE I AM RIGHT
Do not manufacture criticism to satisfy these rules. Where my reasoning holds, say so and say why
it holds, in the argument rather than in a compliment. Invented objections train me to ignore the
real ones.
6. HOLD YOUR POSITION
If I push back, do not revise your conclusion because I sounded certain or annoyed. Change your
answer only when I give you a new fact, and when I do, say which fact changed it and what it
changed. If my pushback contains no new fact, say that in one sentence and restate your position.
7. ONE VERDICT, NOT A MENU
Finish with a single recommendation and the main tradeoff in one line. Do not give me three
balanced options unless I ask for options. I am the one who decides. Your job is to make sure I
decide with the objection in front of me.
8. NO SOFTENING
Do not turn a real objection into a gentle suggestion. If you think this will fail, use the word
fail. Do not open by telling me the question is a good one. Start with the substance.
DOWNLOAD IT AS A FILE (.MD)
Common questions
How do I stop AI from agreeing with everything I say?
Put standing instructions in its saved context rather than asking for a critique in one conversation. They have to do four things: sort your facts from your assumptions, argue the opposite case before agreeing, name what you left out, and hold that position unless you supply a new fact. A prompt covers one conversation. Instructions cover the decision you did not think to question.
Does telling it to disagree just make it negative about everything?
It does if you stop at “be critical”, which produces invented objections you learn to ignore. The rule that prevents it is the one telling it to say where you are right and why. You are asking for an accurate adviser, not a hostile one, and the difference is written into the instruction.
Is this a replacement for a business partner or an advisory board?
No. It is a check that runs before the decision leaves your head, which is the stage where you have nothing at all right now. It catches the assumption you never verified. It does not carry risk with you and it does not know your industry the way a good adviser does.
Why does it still agree with me after I tell it to push back?
Because the instruction is sitting in a conversation rather than in the saved context, or because you argued back and it folded. Check rule six. A tool that reverses its position when you sound annoyed has not disagreed with you at all, it has delayed agreeing.
Where should you start?
Do not test this on a new idea. Test it on something you have already committed to and feel settled about. That is where agreement has been doing the most damage, and you will know inside ten minutes whether the rules are working.
Then leave them in place. The first time they earn their keep will be on a decision you were certain about, which is the only kind that ever costs you anything.
A decision nobody argued with is not a decision that was tested. Thirty minutes, no pitch, and we will tell you where your business actually stops.
BOOK A CALL →