Context Windows: Why AI Sometimes Forgets What You Told It

A context window is the amount of text (measured in "tokens," roughly 4 characters each) that an AI can consider at one time. ChatGPT's context window is around 128,000 tokens. Claude's is 200,000 tokens. If you feed the AI more text than it can hold in its context window, it will lose information or miss important details.

Here's the practical problem: you want to give the AI a detailed client brief, three past proposals for reference, your service packages, pricing, and testimonials—then ask it to generate a perfect proposal. But if all that combined exceeds the context window, the AI will start forgetting or deprioritizing parts of what you gave it.

Signs Your Context Window Is Overloaded

• The AI references early parts of your brief but misses details from the end
• Your service description gets lost halfway through proposal generation
• Testimonials aren't incorporated even though you included them
• The proposal contradicts something you specified in your initial prompt

These aren't AI failures—they're context window limits. The AI isn't ignoring information to be difficult; it's genuinely losing track because you've exceeded its working memory.

How to Work Within the Limits

First, understand your limits. Claude 3 Opus has a 200k token context window—roughly 150,000 words. That sounds huge, but a detailed client brief (2,000 words) + three past proposals (9,000 words) + your methodology doc (3,000 words) + testimonials (2,000 words) = 16,000 words. You're only using 10% of your window, so you're fine.

But if you're trying to include 50 past proposals as reference, you'll hit the limit. That's where you need to be selective.

Second, organize information by priority. Put the most critical information (client brief, specific requirements) first. Reference materials second. Context and background third. This way, even if the AI loses some information, it loses the least important stuff first.

Third, break complex requests into multiple steps. Instead of uploading everything and asking "generate a perfect proposal," you might:

• Step 1: Upload client brief and ask the AI to identify three key goals and two main objections you need to overcome
• Step 2: Then upload your service packages and past proposals, and ask the AI to recommend the best fit and why
• Step 3: Finally, ask it to draft the proposal using those recommendations

This uses the context window more efficiently and produces better output because each step is focused.

When Context Limits Actually Help

Interestingly, context window limits can force discipline. When you can't throw everything at the AI at once, you have to think clearly about what actually matters for this specific proposal. That clarity often produces better results than overwhelming the AI with "everything just in case."

Try this: Next time you ask an AI to draft a proposal, track what information it actually used. Did it reference the testimonials you included? Did it use specific numbers from the client brief? If you notice it consistently missing certain types of information, you're hitting context window limits, and you should break your request into steps or reduce the amount of reference material.