Vibe Code Forum teamAdmin
If your AI tool suddenly "forgets" a decision from an hour ago, or your bill is higher than expected, tokens and context windows are usually why.
Tokens
AI models read and write text in tokens: chunks of characters. In English a token is roughly three quarters of a word, so 1,000 tokens is about 750 words. Code often uses more tokens per line than prose.
Usage is measured in tokens: what you send in (your prompt, the files, the whole conversation so far) plus what comes back. On pay-per-use plans you pay for both; on subscription plans, heavy token use is what makes you hit limits sooner.
The context window
The context window is how many tokens a model can consider at once: your instructions, the conversation, and any files the tool has loaded. Modern models have large windows, but they are not unlimited, and:
- When a conversation gets long, tools drop or summarise older parts. Details from early on can be lost.
- Even within the window, models pay less attention to things buried in the middle of a lot of text.
- Every message re-sends much of the conversation, so long chats cost more per message.
What to do about it
- Start a new chat for each new task. Begin with a short summary: what the project is, what you're doing now, any rules.
- Keep project rules in a rules file your tool reads every time, rather than repeating them in chat.
- Point at specific files instead of "look at everything".
- Write decisions down in a file in the project (for example
NOTES.md): stack choices, what's done, what's next. Tell the AI to read and update it. - Ask for a plan first for big tasks, so the important points are in one short message.
Signs you have hit the limit
The AI repeats mistakes it already fixed, ignores earlier instructions, or contradicts the project's existing style. Don't argue with it; start fresh with a clean summary.
Related: Prompting that works, Keep costs under control.