How to Use AI to Find Why Your Plan Will Fail

Asking an AI "is my plan good?" usually gets you a pep talk. There is a better question to ask first.

Inversion means asking how something would fail instead of how it would succeed. With AI, you ask it to play an adversary and list the exact choices that would guarantee failure, then you fix those first.

The whole idea fits in one line: "How can I fail by doing this?" You paste in your plan, the AI plays your worst enemy, and you get a blueprint for disaster to work through.

Why does AI agree with everything you say?

Chatbots lean towards agreeing with whoever is asking. A 2023 research paper, Towards Understanding Sycophancy in Language Models, found that AI assistants trained with human feedback can favour answers that match what the user already believes over answers that are true.

A question like "Is this architecture good?" invites a yes. So I ask the opposite question.

The prompt

This is the prompt I use. It is built on Charlie Munger's rule of inversion, but he did not write it. Copy it, replace the bracketed part with your own plan, and paste it into any chat AI.

Act as Chalie Munger calls it with inversion principle. You want my plan to fail, and you know exactly how to make that happen.

Here is my plan or decision:
[Paste it here: the goal, the main steps, my constraints, and what I have already decided.]

Do not tell me whether it is a good idea. Do not give me tips for success. Answer these instead:

1. If you wanted to guarantee this fails, what exact choices would you make? Give me the blueprint for disaster, step by step.
2. What are the anti-goals? Which outcomes would look like success but actually be failures?
3. Of everything above, which five failures are most likely given what I have told you? Rank them and say why.
4. For each of those five: what is the earliest warning sign, and what one change would prevent it?

Be specific to my plan. If something is missing, say what you are assuming and carry on.

How to use it

1. Write the plan down. Include the goal, the main steps, your constraints, and what you have already decided. Leave out confidential data.

2. Paste it into the prompt and send it. Do not add "but is it a good idea?" at the end. The prompt tells the AI not to answer that, on purpose.

3. Turn each failure into one of three things. A fix, an accepted risk you have written down, or a reason to drop the plan. What is left is a short checklist, and I tick it off before I commit, not after.

An example you can copy

Here is an example I made up to show the shape of the answer. It is not a real run. Someone wants to quit their job in March and freelance as a bookkeeper, with four months of savings.

Failure	Earliest warning sign	One change
Savings run out before steady clients arrive	The balance drops each month while invoices are still unpaid	Line up some paid work, even part-time, before the leaving date
All the income depends on one client	One client accounts for nearly every invoice	Sign a few smaller clients before leaving, not after
Prices are too low to live on	Busy every day and still short at the end of the month	Work out the lowest rate that covers living costs before quoting anyone

Three places I have used it

These are real. The prompts are paraphrased.

EvoLawyer, a GraphRAG project over Indian Supreme Court judgments. I asked an adversary what choices would guarantee the pipeline fails at scale. It came back with unchecked hallucination on Bare Acts, badly chunked text, keyword search instead of semantic embeddings, and no routing agent. I designed each of those out from day one.

pyMAC, my financial modelling work. I asked for the anti-goals of a data-ingestion pipeline: outcomes that mean it runs every day but is useless. It flagged no normalisation layer, speed chosen over data integrity, and constant manual fixes. I added validation layers before the data reached the models.

A workshop on AI for finance professionals. I asked how to make it a waste of their time. It flagged too much history of language models, generic e-commerce examples, and ignoring data privacy limits. I skipped the history and built it around real finance cases.

The trap: plausible failures that do not apply

The AI can invent believable risks. So each item gets one test before it joins my checklist: can I point to something real that supports it, such as my own records or a person who has done it? If not, I park it.

When not to use it

It shows what could break. It cannot see the future, and the decision stays mine.
It needs a plan or decision I can write down. It is no help for what to have for lunch.
A long list can freeze me. I stop at the failures that are likely and that I can act on.

FAQ

What is the inversion principle?

Inversion means solving a problem backwards: ask how to fail, then avoid that. Charlie Munger made the rule famous. Munger joked that he only wanted to know where he would die, so he would never go there.

Is this the same as a pre-mortem?

Close cousins. Gary Klein described the project pre-mortem in Harvard Business Review in 2007: a team imagines the project has already failed and works out why. This version has the AI play an adversary and hand over the failure blueprint, so one person can run it alone in minutes.

Ask how it fails first

Copy the prompt above and run it on the next decision you are nervous about.