ICM vs Prompt Engineering vs RAG: which is your actual problem?

Ryan “RyMac” McKinney · USMC · 21 August 2026 · 3 minute read

These 3 get argued about as if you have to choose. You do not. They fix 3 different faults, and almost everybody who is frustrated right now is paying to fix the one they do not have.

3 different faults

Prompt engineering is what you say. Wording, examples, asking for the format you want. I split this apart properly in context engineering vs prompt engineering.

RAG stands for retrieval augmented generation. You may hear it thrown around like a product. It simply means going and fetching relevant passages out of a big pile of documents and putting them in front of the AI before it answers.

ICM is arrangement. Where the working files live and what gets read first, so a single agent knows the state of the job without being told again. That is interpretable context methodology.

How to tell which one you have

Your AI has the facts and still answers badly. That is a prompting fault. Fix the wording, give it an example of what good looks like.

Your AI keeps losing what you told it. That is arrangement. No amount of better wording survives the information sliding off the desk, and that sliding has a specific cause.

You need answers out of 4,000 documents you did not write. That is retrieval, and that is a real RAG job. A folder structure will not read a decade of contracts for you.

Why the wrong one gets bought

Because the symptom feels the same from the driver's seat. The AI is unhelpful. That is all you can see.

Underneath, "it does not know our pricing" and "it cannot find the clause in that contract" are not the same fault at all, and the first one is fixed by a folder with your pricing in it, for free, in an afternoon.

They stack, they do not compete

This is the part the arguing hides. Good arrangement makes prompting matter less, because the facts are already present. Good arrangement also makes retrieval work better, because what comes back has somewhere to sit.

The ICM paper is careful about this and it is worth repeating: it does not say the other approaches are wrong. It says they are unnecessary weight on sequential work where a person checks each step.

The cheapest order

1. Arrangement first. It is free and it takes 20 minutes. If it fixes your problem you have spent nothing.

2. Prompting second. Also free, and much easier to judge once the facts are reliably in front of the model.

3. Retrieval last, and only if you genuinely have a document pile problem. It is the expensive one and the one that needs maintaining.

I ran that order myself and it is why I stopped building agents after months of paying for the answer to a question I did not have.

The honest limit

Arrangement will not read your contracts, will not survive being ignored for a month, and will not make the model smarter.

What it does is stop the same fact being re-explained every morning. If that is your actual problem, and for most people running a business it is, start with the template and see whether the other 2 were ever needed.