# Day 04: Why more context can make an AI system worse

When an answer comes back weak, the reflex is to add more context. Bigger windows made that reflex cheap, and it is still often wrong.

> **More context is not the same as better context.**
>
> If the extra text does not change the decision, support a claim, or reduce uncertainty, it is adding cost and noise.

## What actually goes wrong

The model can only use what you give it. It does not follow that everything you give it helps.

Add eight loosely related documents to reach the one that matters and the key fact is now competing for attention with seven thousand tokens of plausible, on-topic, irrelevant material. The evidence is *in* the window and the answer still misses it, which is a genuinely confusing failure, because inspection shows the correct information was right there.

Meanwhile cost rises linearly, latency rises with it, and debugging gets harder because every failure now has a much larger haystack to search.

## The prompt that grows after every bug

There is a specific pattern worth naming. Something fails, so a line gets added to the prompt. Something else fails, another line. Six months later the system prompt is 2,000 tokens of accumulated patches, nobody remembers which ones are load-bearing, and no one will delete anything because no one can prove it is safe to.

At that point the prompt has become an undocumented product specification, and it is one that grows monotonically.

The fix is structural: **separate durable instructions from retrieved evidence.** Instructions are versioned, reviewed, and deliberately curated. Evidence is assembled per request and justified per request. Mixing them is how you end up unable to tell which is which.

## The subtraction test

Most teams test by adding. Test by removing instead.

Cut context until the answer degrades. Whatever you could remove without hurting the result was noise you were paying for on every single call: often a surprisingly large share.

And if you cannot say which piece of context changed the answer, the prompt is carrying weight it has not earned.

A large context window is useful. It is not a substitute for context design. It just raises the ceiling on how much noise you can afford to ship before anyone notices.

*Day 4 of 60 Days of Production AI Systems.*

What are you passing into context because it feels safe, rather than because it has proven useful?

---

**Previous:** [Day 03: Why tokenization quietly affects cost, limits, and reliability](https://blog.vamsiannamreddy.com/day-03-why-tokenization-quietly-affects-cost-limits-and-reliability)  
**Next:** [Day 05: Why vague prompts become vague systems](https://blog.vamsiannamreddy.com/day-05-why-vague-prompts-become-vague-systems)  

*Day 4 of 60 · AI Systems Basics (chapter 1 of 10)*