AI Tools Field Guide

Open edition · no CAC needed

Intermediate — How it works, just enough

Tokens and the context window

Lesson 5 of 17≈4 min

A token is about three quarters of a word, and the context window holds the whole request rather than just your message. When it fills, the product refuses, truncates or compacts, and that is also why a fact in the middle of a long file gets missed.

The cat jumps high. The cat jump s high . 6 tokens
A sentence split into tokens: common words are one token, rarer words several.

What a token is

The model does not read letters or words. It reads tokens: short chunks of text from a fixed vocabulary of a few tens of thousands of pieces. Common English words are usually one token; names, acronyms and numbers are broken into several.

"The" is one token. A hull number or a NAVADMIN reference may be five or six. Code, tables and anything with lots of punctuation cost more tokens per word than prose. The working conversion is simple enough to do in your head.

Text Tokens, roughly
A token About three quarters of a word
1,000 tokens About 750 words, or two typed pages
A ten-page instruction 5,000 to 7,000
A 100-page instruction 50,000 to 70,000
A long spreadsheet exported as text Often more than the document it came from

Tokens are also the unit everything is priced and limited in: the window, the reply length, a thinking budget, the API bill. A spec sheet's "128k" means 128,000 tokens, roughly 96,000 words.

The context window

The context window, one page

All of it, top to bottom, goes to the model as one text. It reads everything and continues from the end.

When the page fills, something has to give. Most tools compact: the oldest conversation and tool results are replaced by a short summary, and detail is lost on someone else's judgment of what mattered.

  1. System rulesThe vendor, sometimes the DepartmentRole, tone, refusals, format and safety rules. You never see them, and they are sent every turn.
  2. AGENT.md, GPT or project instructionsWhoever built the helper you are inThe standing instructions and persona: a form someone filled in, or a markdown file such as AGENT.md.
  3. USER.md and memoryYou, and the harness on your behalfWho you are, your preferences, and saved snippets from earlier chats, re-sent silently.
  4. Skills (.md files)The builder, or youProcedures loaded because the task matched them. The one-line description of every skill is always present; the full text joins when one is used.
  5. Tool definitionsThe harnessA description of every tool the model may call: search, connectors, file reading. They cost tokens whether or not a tool is used.
  6. Tool calls and their resultsThe model and the toolsEvery search it ran and everything the search returned, every file it opened. This is what bloats a long agent session fastest.
  7. Uploaded and project filesYou, or your teamThe documents you attached and the project's files, whole or as retrieved passages.
  8. Conversation so farYou and the modelEvery earlier turn, yours and its, resent in full on every new turn.
  9. Your messageYouThe only part you typed this turn.
  10. Room for the replyReservedThe answer has to fit in the same window, so the harness keeps space for it.
Illustrative shares, not measurements of any one product. The window is not the transcript: it is everything on this page, every turn, plus the room the reply needs.

The context window is the maximum number of tokens the model can attend to in one request: the whole bundle from the previous page, plus room for the reply. It is a hard limit on the model, not a setting you can raise.

Windows differ by model and version, so this guide prints no number for any model: each vendor documents the limit per model; assume tens of thousands of words, not unlimited. Two things about the number are not obvious.

1M tokens: shared window 30k Preamble + tools 400k Document About 570k Chat + reply
One window, shared by the preamble, tools, documents, conversation and the reply.
Not obvious Why it matters
It is shared. Instructions, tool definitions, files, history and the reply all draw on one budget. A window that sounds enormous has lost a chunk to the preamble before you type, and a long conversation eats the rest.
Bigger is not evenly better. A model attends unevenly across a very long context; facts near the start and end are used more reliably than facts buried in the middle. If the answer is on page 60 of a 100-page attachment, ask about page 60, or attach that section alone. Put the instruction that matters most at the start of your message and repeat it at the end.

What happens when it fills

Products handle a full window in one of three ways, and it pays to know which yours does.

Behavior What you see What it costs you
Refuse "This file is too large" or a request that never sends Nothing silently; you split the file and try again
Truncate The oldest turns, or the end of a file, are dropped Silent loss; the model never saw the part that was cut
Compact Older turns are replaced by a short summary and the chat continues Detail lost according to someone else's judgment of what mattered

Compaction is the common case in long chats. The summary keeps the gist and drops the specifics: the exact wording of a correction, the second of three constraints, the fact that you rejected a draft and why.

The model then continues from the summary as if it were the whole history, confident and slightly wrong.

Before: 30 older + 10 recent After: summary + 10 recent
Older turns become a summary; recent turns stay; detail is lost in the exchange.

Six habits that keep the window in hand

Habit Why
One task per chat A chat that has done three jobs carries three jobs' history. Start a new one when the job changes.
Attach the section, not the binder If the answer is in chapter 4, attach chapter 4.
Keep durable material in a project Files and standing instructions in a project are re-supplied cleanly each time instead of accumulating in the history.
Write the brief yourself Before a long chat, and whenever it drifts, put goal, source, decisions and open questions in one message. Your brief beats the product's summary.
Ask what it saw "Which sections of the attachment did you use?" tells you whether the file was read, retrieved from, or truncated.
Watch the reply length A nearly full window leaves no room for a long answer. A reply that stops abruptly is usually the window, not the model.

A staff example

Asked Got Fix
A training officer attaches an entire 200-page curriculum and asks for the prerequisites of one course. Three prerequisites, confidently. Two are right; one belongs to a different course listed later in the file. The same question with the eight relevant pages attached returns the correct three with page numbers. The file was not too big to upload; it was too big to be read evenly.

What to do with this

Estimate the token size of the largest document you use regularly. If it is in the tens of thousands, plan to work from sections. Then read Thinking modes and model tiers, the other dial that trades time and tokens for quality.