My VA kept pasting entire product catalogues into ChatGPT and wondering why it forgot the brand voice
I run a Shopify store selling handmade candles—soy wax, wood wicks, scents like bourbon vanilla and sea salt driftwood. We do about $18,000 a month, mostly direct-to-consumer, with a bump during Black Friday and another in early December when people panic-buy gifts. I have a 1099 contractor, Sarah, who writes product descriptions, collection pages, and Instagram captions. She's good at the work but new to AI. In March I asked her to rewrite 240 product descriptions using ChatGPT because our old ones were flat and didn't mention burn time, wax type, or wick material—details people actually search for.
She came back three days later saying ChatGPT "kept forgetting" the brand voice after the first 30 products. The descriptions started confident and warm, then drifted into generic home-decor language. Words like "elevate your space" and "indulge in luxury" started appearing. I looked at her process and realised she was pasting the entire style guide, all 240 SKU names, and a sample description into one ChatGPT prompt, then asking it to rewrite each product in order. The whole input was over 40,000 words. ChatGPT has a context window—a limit on how much text it can hold in memory at once—and she'd blown past it. By product 31, it had forgotten the style guide. By product 60, it had forgotten the sample.
I needed her to understand context windows, not just follow a workaround. So I used Kryotta's Compare feature to test four models—Claude Sonnet 4.5, Gemini Pro, Llama 3.3 70B, and GPT-OSS 120B—and see which explanation actually made sense to someone who doesn't think about tokens or parameters.
I wrote one prompt asking each model to explain context windows to a non-technical freelancer
I opened Kryotta, turned on Compare, selected all four models, and wrote this:
"Explain context windows to a freelance writer who uses AI for product descriptions. She keeps pasting too much text into one prompt and the AI forgets earlier instructions. Use an analogy that makes sense for someone who isn't technical. Keep it under 150 words."
Compare sent that prompt to all four models at once and showed me the responses side by side. Claude used a notebook analogy: the AI has a fixed number of pages it can look at, and when you add more text, the oldest pages fall off the back. Gemini compared it to a text message thread where old messages scroll out of view. Llama used a shopping cart with a weight limit—add too many items and the earliest ones get pushed out. GPT-OSS went with a whiteboard that gets erased from the left as you write on the right.
I sent all four to Sarah and asked which one made sense. She replied in ten minutes: the shopping cart. She said she could picture adding products to a cart and watching the first ones disappear when it got too full. The notebook and whiteboard felt abstract. The text thread was close, but she doesn't think of old texts as "gone"—she scrolls up and they're still there.
The shopping-cart analogy worked, but only after I added one detail
Sarah understood the concept, but she still wasn't sure how much text was too much. She asked if there was a number. I went back to Compare and added a follow-up:
"How many words can Claude, Gemini, Llama, and GPT-OSS hold in one conversation before they start forgetting? Give a rough word count a freelancer can use as a rule of thumb."
Claude said about 75,000 words for Sonnet 4.5. Gemini said 90,000 for Pro. Llama estimated 50,000 for the 70B model. GPT-OSS said 80,000 for the 120B version. All four added that the number includes both what you write and what the model writes back, so a long conversation eats into the window faster than a single prompt.
I told Sarah to assume 40,000 words as a safe ceiling—half of the lowest estimate—and to count her input before pasting. Her style guide was 1,200 words. Each product description draft was about 80 words. That meant she could do roughly 480 products in one session if she pasted nothing else, or about 120 products if she included examples and the style guide. She started batching her work into sets of 100 products with the style guide pasted at the top of each batch. The descriptions stayed consistent.
I tested whether the analogy held up when I asked each model to explain why windows exist
I wanted to know if Sarah would hit a wall later, so I asked Compare:
"Why do AI models have context windows at all? Why not just remember everything? Explain it in one sentence to someone who isn't a developer."
Claude said it's because reading and holding more text in memory costs more computing power, and there's a practical limit. Gemini said models process text in chunks, and bigger chunks slow everything down. Llama said memory isn't infinite and longer context makes the model slower and more expensive to run. GPT-OSS said the model has to "pay attention" to every word in the window, and paying attention to a million words at once isn't possible yet.
All four were accurate, but Claude's answer was the one Sarah would remember. She doesn't need to know about attention mechanisms. She needs to know that more text = more cost and slower responses, so there's a trade-off. I sent her Claude's explanation and added, "This is why I don't want you pasting the entire product catalogue into one prompt—it's not just that it forgets, it's that it's wasteful."
The wrong analogy led to 40 descriptions that had to be rewritten
Before I ran the Compare test, I'd tried explaining context windows to Sarah myself using a dinner-party analogy: the AI can only remember the last hour of conversation, so if you talk for three hours, it forgets what you said at the start. She nodded, said she understood, and then went back to pasting 60 products at a time with no style guide because she thought the model would "remember the tone from earlier in the conversation."
It didn't. I got 40 descriptions back that were fine for the first 20 products, then drifted into stiff, generic language for the next 20. She'd assumed the model had absorbed the brand voice from the earlier batch and didn't need the style guide repeated. The dinner-party analogy made her think the AI had short-term memory that faded gradually, like a person forgetting details. The shopping-cart analogy made her think of a hard cutoff—once the cart is full, the first items are gone, not just fuzzy. That mental model was more accurate and led to better habits.
I had to pay her for the 40 rewrites. That cost me $320 and three days. The Compare test took ten minutes and cost nothing.
I used Compare again to see which model explained tokens vs words most clearly
Sarah asked a follow-up question: "You said 40,000 words, but I've seen people online talk about tokens. Are those the same thing?" I didn't want to guess, so I ran another Compare prompt:
"Explain the difference between tokens and words to a freelancer who's trying to figure out how much text fits in a context window. One short paragraph."
Claude said a token is usually a piece of a word—about three-quarters of a word on average in English—so 100 tokens is roughly 75 words. Gemini said tokens are chunks of text the model reads, and common words like "the" are one token, but longer words like "strawberry" might be two or three. Llama said most English words are one or two tokens, so you can estimate by dividing your word count by 1.3. GPT-OSS said a token is the model's "unit of reading" and you can think of it as a syllable or two.
I sent Sarah the Claude and Llama explanations. She said Llama's formula—divide by 1.3—was the easiest to use. Now when she's not sure, she takes her word count, divides by 1.3, and checks it against the token count Kryotta shows in the bottom corner of the prompt box. She's never been off by more than 10%.
The Compare feature saved me from testing four models over four weeks with four sets of real work
If I'd tested these explanations one at a time, I'd have sent Sarah the Claude explanation, waited to see if she got it, then tried Gemini a week later when the next batch of product descriptions came in, then Llama, then GPT-OSS. By the time I'd cycled through all four, it would be June and I'd have a dozen batches of inconsistent descriptions to clean up.
Compare let me test all four in one session and pick the best answer in ten minutes. I didn't have to guess which model would work. I didn't have to risk Sarah misunderstanding and wasting another three days. I just sent the same question to four models, showed her the results, and asked which one clicked. The shopping-cart analogy from Llama won, and her next 200 descriptions were consistent from start to finish.
I still use Claude for the actual product-description writing—it's better at brand voice—but I use Llama through Compare whenever I need to explain something technical to a contractor. It's good at analogies that don't assume the reader knows how software works.
Questions people ask
Can I use Compare to test models on other tasks, or is it just for explanations?
You can use it for anything—writing, analysis, summaries, code. I've used it to test which model writes better Instagram captions for candle launches and which one structures email sequences more clearly. It's faster than switching between models manually.
Do all four models charge the same, or does using Compare cost more?
You pay for what you use. If you send one prompt to four models, you're billed for four responses. Kryotta shows the cost per model before you send, so you can skip the expensive ones if you're just testing. I usually compare Claude, Gemini, and Llama and leave out GPT-OSS unless I need a fifth opinion.
How do I know when my VA is hitting the context window limit?
The descriptions start drifting in tone or the model starts repeating itself. If the last 50 products sound different from the first 50, she's probably pasting too much. I taught Sarah to check the token count in Kryotta before sending and to stay under 30,000 tokens per batch.
Is 40,000 words a safe rule for every model, or does it change?
It changes. Claude Sonnet 4.5 handles more than Llama 3.3 70B. I told Sarah 40,000 words because it's conservative and works across all the models she's likely to use. If she sticks to Claude, she could go higher, but I'd rather she stay cautious than have to redo work.
Compare is one of those features I didn't think I needed until I had a contractor asking questions I couldn't answer off the top of my head. Now I use it every time I'm not sure which model will do the job better. You can try it at app.kryotta.ai/auth—it's faster than testing models one at a time and cheaper than paying for rewrites.
