Let me tell you something that changed how I use AI.
AI does not read your words.
Not the way you do. Not the way your friend does. Not the way your teacher or your boss reads what you write. When you type a sentence into ChatGPT or Claude or any AI tool, something happens before the AI even begins to think about your question. It chops your words into pieces.
Those pieces are called tokens. And once you understand what that means, you will never write a prompt the same way again.
Start here. Forget AI for a second.
Think about how you learned to read.
You did not start with sentences. You started with letters. A. B. C. Then you learned that letters form syllables. Then syllables form words. Then words form sentences. Your brain learned to look at “communication” and instantly see one word. But a child learning to read might see “com-mu-ni-ca-tion” as five separate pieces.
AI is like that child.
When you type “I love Lagos” into an AI tool, it does not see three words the way you do. It breaks them into pieces called tokens. “I” might be one token. “love” might be one token. But “Lagos” might become two tokens: “Lag” and “os”. The AI learned “Lag” and “os” as separate pieces because it did not see the word “Lagos” often enough in its training to learn it as one whole unit.
Common English words like “the”, “is”, “and”, “what” get their own single token. They are efficient. The AI has seen them millions of times. But less common words, names, slang, and words from other languages get chopped into smaller pieces. More pieces means more work for the AI. And more work means worse results and higher cost.
The translator analogy
Here is another way to think about it.
Imagine you hire a translator who charges by the word. You give them a letter to translate. In English, the letter is 200 words. Simple, fast, affordable.
Now imagine you give the same translator the same letter written in Yoruba. The translator does not speak Yoruba fluently, so for every Yoruba word, they need to break it into smaller parts, look up each part, figure out how the parts connect, and then understand what you meant. The same letter that was 200 words in English now takes the translator 500 “word-units” of effort in Yoruba. Same meaning. Same letter. But it costs more and takes longer because the translator was not trained in Yoruba.
That is exactly what happens inside AI.
The bucket: how to think about token limits
Here is the most important thing to understand. Every time you have a conversation with AI, you are sharing a bucket. Your prompt, the AI’s thinking, and the AI’s response all come out of the same bucket. One budget. Shared.
Who decides how big the bucket is? The company that built the model. Anthropic decides Claude’s bucket holds 200,000 tokens. OpenAI decides GPT-4’s bucket holds 128,000 tokens. Google decides Gemini’s bucket holds over 1,000,000 tokens. It is a design decision, like how a car manufacturer decides the size of the fuel tank. You do not choose the bucket size. You choose which model to use, and you work within its limits.
Is that per message or per conversation? Per conversation. If a model has a 128,000 token bucket, that is 128,000 tokens total for everything: every message you have sent, every response the AI has given, every document you pasted in, and all of the AI’s thinking. Not per message. Total. The entire conversation shares one bucket.
What happens when the bucket is full? The AI does not crash. It does not stop. It starts forgetting. It quietly drops the oldest parts of the conversation. So if you have been chatting for a while and the bucket fills up, the AI literally loses what you said at the beginning. It can still see your latest message, but the early context is gone. That is why long conversations sometimes feel like the AI “lost the plot.” It did. The oldest messages fell out of the bucket to make room for new ones.
How big is a token? A rough guide: in English, one word is about one token. Some longer or unusual words might be two tokens, but for everyday counting, one word equals one token is close enough. So 100 words is roughly 100 tokens. 1,000 words is roughly 1,000 tokens, which is about two pages of text. 10,000 tokens is about 20 pages.
How big is a typical conversation?
This is where it gets practical. Here are some real examples so you can start estimating in your head.
Small request (100 to 300 tokens total) “Write me a short email to my landlord asking for a repair.” Your prompt is maybe 15 tokens. The AI’s response is maybe 150 tokens. Total: around 165 tokens. Quick, cheap, easy. This is like asking someone to pass you a glass of water. The bucket barely notices.
Medium request (500 to 2,000 tokens total) “Find me the closest amala buka to Yaba and tell me their prices and opening hours.” Your prompt is maybe 20 tokens. The AI might search, think about what it found, and write a detailed response. Total: maybe 800 tokens. Still comfortable. Like filling a small bowl from the bucket.
Large request (3,000 to 10,000 tokens total) “Here is my cover letter and the job description. Tell me what is wrong with my letter, score it, and rewrite the weak paragraphs.” You are pasting in maybe 500 tokens of text. The AI reads it, analyses it, and writes a detailed response. Total: maybe 4,000 tokens. Now you are using a meaningful chunk of the bucket.
Very large request (10,000 to 50,000 tokens total) “Here is a 30-page business plan. Summarise it, identify the five weakest sections, and rewrite the executive summary.” The document alone might be 10,000 tokens. The AI’s analysis and rewrite adds another 5,000. Total: maybe 15,000 tokens. You are now pouring a lot into the bucket.
Massive request (50,000+ tokens) “Here is the full text of a 200-page book. Analyse the main arguments, compare them to these three other sources, and write a critical review.” This could easily hit 80,000 to 100,000 tokens. Only the largest AI models can handle this at all. And even then, the quality of the AI’s thinking may drop because the bucket is stretched so thin.
The pattern: the bigger your input, the less room the AI has to think carefully and respond well. And if your conversation keeps going and the bucket fills up, the AI starts dropping your earliest messages to make room. That is why a long conversation about a business plan might end with the AI forgetting the details you gave it an hour ago. The old messages literally fell out of the bucket.
The bucket is the entire conversation, not just one message
This is something most people do not realise.
The bucket is not just your last message and the AI’s last response. It is the entire conversation. Every message you have sent. Every response the AI has given. Every back and forth. All of it is sitting in the bucket at the same time.
If you look at the left side of ChatGPT or Claude, you will see a list of conversations. Each one of those is a separate bucket. When you open one and keep adding messages, you are pouring more and more into that same bucket. The AI is re-reading the entire conversation every time you send a new message. That is how it remembers what you said earlier. It is not actually remembering. It is re-reading everything from the beginning.
So a conversation that started with “write me an email” and then continued with “now make it more formal” and then “add a paragraph about the budget” and then twenty more messages, that conversation is using tokens for every single message, yours and the AI’s, all stacked up.
This is why AI tools eventually tell you to start a new conversation. You have filled the bucket. There is no more room. The AI can no longer read the full conversation and still have space to respond properly.
This is also why your AI responses sometimes get worse the longer a conversation goes on. The bucket is getting full. The AI is trying to re-read a long history and still give you a good answer with whatever space is left. The quality drops because the space is shrinking.
The practical lesson: start a new conversation when you switch to a new topic. Do not use one long conversation for everything. Each new conversation gives you a fresh empty bucket. Your prompts will be more efficient and the AI’s responses will be sharper.
If you are working on a big project, break it into separate conversations. One for research. One for writing. One for editing. Each conversation stays focused and the bucket stays clean.
What about when AI “thinks”?
Some newer AI models have a feature where they think before they respond. You might have seen this with ChatGPT’s reasoning mode or Claude when it works through a problem step by step. That thinking uses tokens too. It all comes from the same bucket.
So when you ask AI to “think step by step” or “reason carefully,” you are giving it permission to use more of the bucket on thinking, which means a potentially better answer but less room for a long response. This is usually worth it. A short, well-reasoned answer beats a long, shallow one.
Why African languages cost more
This is the part that should make you pay attention.
When you type in English, common words are one token each. Efficient. When you type in Yoruba, each word might become three or four tokens because the AI’s dictionary was not built on much Yoruba text. That means the same question costs 2 to 3 times more tokens in Yoruba than in English. Your question takes up more of the bucket. The AI has less space to think and respond.
And because the AI saw less Yoruba in its training data, it understands the chopped-up tokens less well too. It is like the translator trying to read a language they barely speak. They burn more effort and still get it wrong sometimes.
This is not a bug. It is a structural disadvantage built into how these models were created. English speakers get cheaper, faster, better AI. Everyone else pays a tax. The same applies to Pidgin English, Igbo, Hausa, and most African languages. Even Nigerian English with local expressions and slang is less efficient than standard English.
This is one of the reasons why learning to prompt well matters so much. Clean, precise English prompts make the most of your tokens. You get better results for the same cost.
How to count tokens in your head
You do not need to be exact. The rough rule is enough: one English word is about one token.
If you write a 20-word prompt and the AI writes 300 words back, that conversation used about 320 tokens. If you paste a 2,000-word document and ask for a 500-word summary, that is about 2,500 tokens.
When you are deciding how to write a prompt, ask yourself: how big is this task?
If you are asking for a short email, keep your prompt short. You do not need to give the AI your life story.
If you are asking the AI to analyse a long document, keep your instructions tight and specific. “Summarise this in 200 words” is better than “Tell me everything you think about this document and give me all your thoughts and recommendations and any other ideas you might have.” The first instruction tells the AI exactly how much of the bucket to use on the response. The second gives it no guidance, so it might ramble until it runs out of space.
If you are asking for code, know that code is token-heavy. A single function might be 200 tokens. A full application could be 5,000 to 20,000 tokens. When you ask AI to build something, be specific about what you want so it does not waste tokens generating code you do not need.
The bottom line
AI reads by chopping your words into pieces. Everything you type, everything the AI thinks, and everything it writes back comes from the same shared budget. The cleaner your words, the more room the AI has to give you something useful.
This is not a technical detail for programmers. It is the most practical thing you can learn about AI. Because once you understand that you are sharing a bucket with the AI, you stop pouring in filler and start pouring in instructions.
Write short. Write clear. Tell the AI exactly what you want, how much you want, and in what format. Do not waste words on pleasantries or vague instructions. Every unnecessary word is space the AI could have used to give you a better answer.
This is what we mean at Kade Labs when we say prompting is a skill. It is not about tricks or hacks. It is about understanding what is happening on the other side of the screen, and using that understanding to get more from every interaction.
Try this right now
Open any AI tool. ChatGPT, Claude, Gemini, any of them.
Ask it something simple: “Write me a professional email asking my boss for two days off next week for a family event.”
Now ask the same thing with a longer, less focused prompt: “Hi, I was hoping you could help me with something. I need to write an email to my boss and I want it to sound professional. The thing is I need to take some time off work, probably like two days, next week, because I have a family event coming up and I really need to be there. Can you write something that sounds good and not too casual but also not too formal?”
Compare the results. The first prompt will give you a tighter, better email. Not because the AI likes short messages. Because you gave it more room to work.
Now imagine the difference across hundreds of prompts. Across a career. Across a business built on AI tools.
That is why this matters.
Kade Labs teaches young Nigerians and Africans to use AI, build with AI, and earn from AI. 25 weeks, 4 stages, 4 certificates. Join the community for ₦1,000 a month.