Context Engineering vs Prompt Engineering: The Actual Difference.

Ryan “RyMac” McKinney · USMC · 15 August 2026 · 5 minute read

Prompt engineering is what you say to the AI. Context engineering is what the AI can see while you say it. That is the distinction. Once you have it, most of the advice you were given sorts itself into useful and useless.

The 1 sentence version

A prompt is the question. Context is everything on the desk when the question gets asked.

The same request, both ways

Prompt engineering approach: you spend 20 minutes crafting "You are an expert proposal writer with 15 years of experience. Write in a professional but warm tone. Use the following structure..."

Context engineering approach: you point it at the folder. That folder holds your last 5 proposals. It holds your pricing. It holds the call notes from this client. Then you say "write the proposal."

The second one wins, and it is not close.

Why prompting used to matter more

It genuinely did, and I do not want to rewrite history here.

Earlier models needed steering. Phrasing changed results a lot, so the skill worth having was phrasing.

Models got better at understanding intent. What they cannot do, and will never do, is know something you never showed them. So the bottleneck moved from how you ask to what is available when you ask.

Is prompt engineering dead?

No, and anybody telling you it is has something to sell.

Prompting still matters for shaping output: format, length, what to leave out, what to do when uncertain. That is real and it is worth learning.

It just stopped being the lever that fixes "it does not know my business." No prompt fixes that. Only material does.

How to tell which problem you have

Look at what went wrong.

Wrong shape, right facts? Prompt problem. It knew the situation and formatted it badly. Fix the ask.

Right shape, wrong facts? Context problem. It wrote a beautiful proposal with the wrong price in it. No amount of rephrasing fixes that.

Almost everything people complain about is the second one.

What fixing the second one looks like

~/business/
├── CLAUDE.md          <- the map. 11 lines. always read first
├── Context.md         <- the desk. what we're doing today
└── ops/
    ├── clients/
    ├── offers/
    ├── sops/
    └── archive/

1 file it always reads first. 1 file that says what we are doing. Folders with the actual material.

The method behind it is Interpretable Context Methodology, from the paper by Jake Van Clief and David McDermott. The full breakdown: Interpretable Context Methodology in plain English. Jake teaches it himself over at Clief Notes.

The wider concept is in the context engineering post, and the reason it forgets mid-session is the context window.

Every ICM Folder System I have built, the free ones and the ones people paid for, is on 1 page. Take the 1 that matches where you are.