Picture a single sheet of paper.
When you write a prompt, the AI fills that sheet. As long as your instructions fit cleanly on that page, the model reads every word carefully. Clear context. High accuracy. Good output.
Now imagine cramming ten pages worth of instructions onto that one sheet. Smaller font. Denser text. Walls of context. The model still reads it — but it can't give every word equal attention. It starts skimming. It misses details. It loses track of what you actually wanted.
This is the context window problem most people understand. But there's a second problem that almost nobody talks about.
Even within a normal-length prompt, AI doesn't pay equal attention to everything you write.
The Primacy/Recency Effect
The primacy/recency effect is a well-documented phenomenon in human psychology. When people are asked to recall a list of items, they remember the ones at the beginning and the end much better than the ones in the middle. The middle fades.
AI language models have the same pattern.
I was teaching a session with Blake Eastman, and we got into context window mechanics. The insight that came up: AI models selectively attend to different parts of your prompt. The beginning gets high attention. The end gets high attention. The middle? The model skims.
This means your prompt has three zones, and they're not equal.
Zone 1 — The top: High attention. The model processes this carefully. Whatever you put here gets the most focus.
Zone 2 — The middle: Lower attention. Background context, supporting details, examples. Fine for supporting information, but the model's focus has drifted by now.
Zone 3 — The bottom: High attention again. Output requirements, format expectations, what to avoid. The model re-engages here because it's processing the final instructions before responding.
Why This Explains a Lot of Failures
Think about how most people structure prompts.
A common pattern: write a long setup paragraph explaining the background, then the actual task buried somewhere in paragraph three, then a few format preferences at the end.
The background gets read. The format at the end gets read. The actual task in paragraph three? Partially skimmed.
This is why you'll sometimes get AI outputs that technically respond to your prompt but miss the specific thing you asked for. The instructions were there — they were just in the middle.
How to Restructure Your Prompts
The fix is straightforward once you understand the zones.
Start with what matters most. Your role definition, the primary task, and the most critical constraints go at the very top. Don't warm up with background. Get to the point first.
Put supporting context in the middle. Background information, examples, data, and reference material belong here. The model needs it, but it doesn't need to process it with maximum attention.
Close with output requirements. Format, length, tone, what to avoid — these go at the bottom where the model re-engages. This is the last thing it processes before generating a response, so it's the best place for precision requirements.
If you use a framework like OCE — Outcome, Context, Expectations — this maps cleanly. Outcome at the top. Context in the middle. Expectations at the bottom. The structure isn't just logical, it's engineered to match where the model's attention is highest.
The Memo Analogy
Think of a well-written business memo.
It leads with the bottom line. The ask is in the first sentence, not buried in paragraph four. Then the supporting information and context. Then the action items or next steps at the close.
That structure exists for a reason — it matches how busy people actually read. They scan the beginning to decide if it's worth reading, and the end to see what's being asked of them. The middle is where the details live.
Prompts work the same way. The model is a busy reader. Structure for how it actually reads, not how you'd write an essay.
One Quick Test
Pull up a prompt you've been struggling with. A prompt that technically includes everything the AI needs but still produces mediocre outputs.
Read it and highlight the single most important instruction. Where does it live?
If it's in the middle, move it. Put it in the first two sentences. Then rerun the prompt.
A lot of the time, that one change is enough.
Thanh Pham is the founder of Asian Efficiency and runs AI workshops in Austin. To improve your AI prompting skills from the ground up, check out the 4-Day AI Sprint.
