Lesson 1.2: Tokens, context and why wording matters
Learn what tokens and context windows are, how to budget the text you send, and how word choice and order quietly change what a model gives back.
By the end of this lesson you will be able to estimate how many tokens a piece of text uses, work out whether your prompt and attachments fit in a model's context window, explain why very long chats and very long documents behave worse than short ones, and choose words that remove ambiguity from a request. You will use a budget calculation on a real-sized example, see why a leading question gets a flattering answer, and learn a small set of word swaps that make prompts sharper. You should have read Lesson 1.1, since this lesson builds on the idea that a prompt is everything the model reads. You need no technical background, and the numbers here are deliberately rough, because they differ between models and languages.
When you type into an assistant, the text feels continuous. The model does not see it that way. It sees pieces, and it can only hold a limited number of them in view at once. Those two facts, how text is cut up and how much fits, explain a large share of everyday puzzles: why a long document gets summarised badly, why an assistant "forgets" what you said earlier, and why a small change of wording changes the answer.
What a token is
A token is a small chunk of text, the unit in which a model reads and writes. Common short words are often a single token. Longer or rarer words are split into several pieces. As an illustration, a word such as "unbelievable" might be handled as pieces like "un", "believ" and "able", though the exact split differs from one model to another. Punctuation, spaces and numbers count too.
For ordinary English, two rules of thumb are good enough for planning.
One token is roughly four characters, or about three-quarters of a word.
So 100 tokens is about 75 words, and 1,000 tokens is about 750 words, a page and a half of text.
To turn words into tokens, divide by 0.75, which is the same as multiplying by 4 and dividing by 3. A 150-word email is about 200 tokens. A 1,000-word memo is about 1,330 tokens.
Text · Words · Tokens (approximately)A short email · 150 · 200
A two-page memo · 1,000 · 1,330
A 10-page report · 5,000 · 6,670
A 50-page contract · 25,000 · 33,330
A short novel · 75,000 · 100,000
These are estimates. Text in other languages, including many Indian languages, usually takes more tokens per word than English, and so do code, tables full of numbers, and long strings such as account numbers or web addresses. If your work involves those, expect your real count to be higher, and measure with your tool's counter where it offers one.
You can try the idea directly with the widget below. Paste or choose text, and watch the count against a window size.
What a context window is
The context window is the maximum number of tokens the model can consider at one time. Everything counts against it: your standing instructions, the whole conversation so far, any attached document, and the reply the model is writing. Think of it as the size of the desk. Everything the model uses has to be on the desk at once, and the desk has a fixed size.
Windows have grown over the years, from a few thousand tokens to hundreds of thousands in some tools, but they are never unlimited, and bigger is not always better, as we will see. What happens when you exceed the window depends on the tool. Some stop with an error. Some quietly drop the oldest part of the conversation. Some cut your attachment short. The dangerous cases are the quiet ones, because the reply looks normal while parts of your input were never read.
Predict first: you attach a 6,000-word report, paste in a 150-word instruction, have a four-turn chat history of about 1,200 words, and expect a 500-word answer. The tool has an 8,000-token window. Will it fit? Work it out before reading on.
Add the words: 6,000 + 150 + 1,200 + 500 = 7,850 words. Divide by 0.75 to get tokens: about 10,470. That is well over 8,000, so it does not fit, by roughly 2,470 tokens. The document alone is about 8,000 tokens, which fills the whole window before anything else is counted. Many people are surprised by this: a document that looks modest as a file can take the entire context.
The chart shows where to cut. The instruction you typed is tiny. The document dominates. If it does not fit, the right move is to send only the relevant sections, not to shorten your instruction.
Bigger windows are not a free lunch
A very large window sounds like the end of the problem: just paste everything. Three limits remain.
Attention is uneven. Studies of long inputs have reported that models tend to use information near the beginning and the end more reliably than information buried in the middle. How strong this is varies by model and keeps changing, but the safe habit is clear: put the most important material and the instruction where they are hard to miss, and test your own tool on a long input before you trust it.
More text means more noise. Every extra page is another place for the model to get distracted. A tight, relevant excerpt often gives a better answer than the whole file.
Cost and speed. Most paid tools charge by the token, and longer prompts take longer to process. For occasional personal use the cost is small. For a team that sends thousands of requests a day, a few hundred extra tokens in every prompt is a real bill.
With three examples of 150 tokens each, sent 2,000 times a day, the extra is 900,000 tokens, or 0.9 million. At an example price of 200 per million that is 180 a day, or 5,400 a month, for text that may add little after the first example. This is not a reason to skip examples, which Lesson 2.1 shows are valuable. It is a reason to measure what they buy.
Long conversations drift
A chat is just a longer prompt. Each turn, the tool sends the whole conversation so far, plus your new message. As the chat grows, three things happen. Your original instructions become a smaller part of what the model reads. Contradictions pile up, such as "make it formal" ten turns ago and "make it friendlier" five turns ago. And the conversation may hit the window, at which point the earliest turns, which often held your most important instructions, can be dropped.
The practical fixes are simple. Restate the key rules when a chat gets long. Start a fresh chat for a new task, pasting in only what the new task needs. And when a conversation has produced something good, ask for a summary of the decisions made and paste that into a new chat as the starting context.
Why wording matters
Because the model predicts what is likely to follow your words, the words you pick steer the result. This is not magic. It is the same fact seen from the other side: the prompt is the only evidence about what you want. Four patterns come up constantly.
Ambiguous words
Words like "short", "recent", "simple", "professional", "best" and "good" mean different things to different people, and the model will choose its own meaning. Replace them with measurable ones.
Weak wording · Why it is weak · Sharper wording"Make it short" · Short is 3 lines or 3 pages? · "Under 100 words"
"Recent data" · Last week or last year? · "Data from January to September 2026"
"Simple language" · Simple for whom? · "Explain it to a first-year student with no finance background"
"Professional tone" · Formal, warm, cold? · "Polite and direct, like an email to a senior client"
"Best options" · Best by what? · "The three cheapest options that arrive within 3 days"
"Summarise" · How, for whom? · "Summarise in five bullets for a busy director"
### Action verbs
The verb tells the model what kind of work to do. "Summarise", "rewrite", "critique", "extract", "compare", "brainstorm" and "translate" each produce a different kind of output. If you ask to "look at" a draft, you may get a summary when you wanted fixes. Pick the verb that names the deliverable: "Rewrite this paragraph to be half as long", not "Look at this paragraph".
Leading questions and flattery
Models are trained to be helpful and agreeable, and they often follow the lean of a question. Ask "This plan is good, right?" and you will usually be told it is. Ask "What are the weaknesses of this plan?" and you will usually get weaknesses. Neither is a fair test on its own. For an honest read, ask for both sides in the same prompt, and say you want candour.
text
Review the plan below. List its three strongest points and its three biggest
risks, in that order. Do not soften the risks. If a risk would make the plan
fail, say so plainly.
Order and emphasis
Position matters. The same list of options can come back differently ranked if you reorder it, because earlier items are weighed differently. When you give several items, say whether the order is meaningful. And when a single rule is critical, such as "never state a price that is not in the attached sheet", state it clearly and, in long prompts, repeat it near the end, where it is least likely to be lost.
A related habit: say what to do, not only what to avoid. "Do not use jargon" tells the model what to leave out but not what to put in. "Use plain words a school student would know" gives it something to aim at.
Why your token count may be higher than the rule of thumb
The three-quarters-of-a-word rule was built on English prose, and it flatters other kinds of text. If your work involves any of the following, plan for a larger count.
Languages other than English. Many models split Hindi, Tamil, Bengali and other scripts into more pieces per word than English, because less of that text was in the data the model's vocabulary was built from. A passage that is 200 tokens in English can plausibly be two or three times that in another script. The multiple varies by model, so test with your tool before you plan a budget.
Numbers and tables. A long number such as an account or invoice number is often broken into several tokens, and a spreadsheet pasted as text carries many separators.
Code and structured data. Brackets, indentation and variable names add up quickly.
Mixed-language text. Writing in English with a few words of another language can cost more than either alone.
A practical habit is to measure once. Take a typical document you work with, find its token count in your tool or a counter, and divide by its words. That ratio, perhaps 1.3 for English and 2.5 for a regional-language document, becomes your own planning number and is more reliable than any rule of thumb. It also tells you when a task that looks small in words is large in tokens.
Carrying work forward without carrying the whole chat
Because a long chat both dilutes your instructions and fills the window, a good professional habit is to checkpoint. When a conversation has reached a useful point, ask for a compact record of where you stand, and use it to start fresh. A prompt such as the following works with any assistant.
text
Summarise this conversation as a briefing I can paste into a new chat.
Include: the goal, the decisions we made and why, any facts or numbers I gave
you, the style rules I asked for, and what is still open. Keep it under 250
words and do not add anything we did not discuss.
Say a drafting session reached 6,000 words of back-and-forth, about 8,000 tokens. The 250-word checkpoint is about 330 tokens, a reduction of roughly 96 per cent, and it holds the parts that matter: the decisions, the facts and the style rules. Pasted into a new chat, it leaves almost the whole window free, and your standing rules are once again a large share of what the model reads. Read the checkpoint before you use it, because summaries can drop details or introduce small errors. If a number matters, check it against your own notes.
Worked example: fixing a prompt that was too long and too vague
A finance analyst pastes a 9,000-word policy document and types: "Summarise this and tell me what is important." The tool has a window of 8,000 tokens. The document alone is 12,000 tokens, so the tool silently keeps only the first part. The summary describes the introduction in detail and says nothing about the later sections that matter.
The fix has three parts. First, cut the input: the analyst pastes only the two sections that apply to a loan approval, about 1,800 words or 2,400 tokens. Second, say who and why: "I am a credit analyst checking whether our policy allows lending to self-employed applicants." Third, name the deliverable and the limits: "List each rule that applies, quote the sentence it comes from, and write 'not stated' if the sections do not cover something." The prompt now uses about 3,000 tokens in all, fits comfortably, and carries a clear task. The result is a short list of rules with quotations that the analyst can verify, in place of an essay about the introduction.
Common mistakes
Pasting a huge document without checking the fit. Estimate the tokens first. If it does not fit, send the relevant parts.
Assuming a big window means perfect recall. Test whether your tool uses the middle of a long input.
Continuing one long chat for many tasks. Start a fresh chat and paste in only what you need.
Using words like "short" and "professional" without measuring them. Give numbers, an audience and an example.
Asking leading questions. Request both strengths and weaknesses.
Forgetting that the answer uses tokens too. Leave room for the reply.
Trusting the count blindly. Counts are estimates, and other languages and tables run higher.
Key takeaways
A token is a small chunk of text, roughly four characters or three-quarters of a word. Divide words by 0.75 to estimate tokens.
The context window is the total the model can hold at once, and it includes your instructions, the history, attachments and the reply.
An attached document often takes most of the window. Trim it to the relevant parts when you can.
Bigger windows help but attention can be uneven, more text adds noise, and longer prompts cost more at volume.
Long chats drift. Restate key rules, start fresh for new tasks, and carry forward a summary.
Wording steers the result: replace vague words with measurable ones, choose the verb that names the deliverable, avoid leading questions and say what to do.
Check yourself
1. Roughly how many tokens is a 3,000-word document?
3,000 ÷ 0.75 = 4,000 tokens.
2. A tool has a 4,000-token window. You send a 1,500-word document, a 100-word instruction, and expect a 400-word answer. Does it fit?
2,000 words ÷ 0.75 = about 2,670 tokens, so yes, with about 1,330 to spare.
3. Why might a very long chat give worse answers than a fresh one?
Early instructions get diluted or dropped, contradictions build up, and the window may fill.
4. Replace "Write a short, professional summary" with a sharper version.
For example: "Write a summary of under 100 words for a senior client, polite and direct, in three bullets."
5. Why is "This plan is solid, isn't it?" a poor way to get a review?
It leans the answer towards agreement. Ask for strengths and risks together and ask for candour.
Try this
You need an assistant and a long document you can safely use, such as a public report. Set aside forty minutes.
1. Estimate the document's tokens from its word count. Compare your estimate with the tool's counter or the widget above if you can.
2. Add the instruction, a rough answer length and any earlier conversation, and use the budget simulator to see how much of a window you would use.
3. Test position. Paste the same long text twice in two fresh chats. In the first, put your question at the top. In the second, put it at the end. Ask something whose answer sits in the middle of the document. Compare the answers.
4. Test wording. Ask for the same task three ways: "Summarise this", "Summarise this in five bullets for a busy director", and a version with an audience, length, focus and one exclusion. Score each against this rubric, one point per yes.
Question · Version 1 · Version 2 · Version 3Right length? · · ·
Right audience? · · ·
Focused on what matters? · · ·
Free of invented facts? · · ·
5. Test the leading question. Ask "Is this plan good?" about a plan of your own, then ask for strengths and risks. Note how the answers differ.
6. Write down the three sharper words you will use from now on instead of vague ones.
A good result is a short note with your token estimate, which position gave the better answer on your tool, and a before-and-after pair of prompts. In the next lesson you put these parts together into a structure you can reuse: the anatomy of a good prompt.
Tags: Prompt Engineering Tokens Context window Wording Beginner