Understanding Tokens and Context
When you use generative AI, the model can only work with a limited amount of information at a time. Understanding that limit can help you choose the right model, write better prompts, and avoid using more computing resources than you need.
What is a token?
AI models process text in small pieces called tokens. As a rough rule of thumb, one token is about 3/4’s of an English word, although the actual number varies depending on the language and content. It’s important to note that your prompt is not necessarily the only information the model processes. Depending on the AI tool, its available context may also include previous messages in the conversation, instructions, and relevant content from files, other chats, or other sources.
When choosing an AI model, use the lightest model and reasoning level that can reliably handle the task. More capable models and more intensive reasoning can improve difficult answers, but they can also require more computation and take longer. Save those capabilities for work that benefits from them.
Make your AI usage more efficient
A few habits can help you get better results without sending unnecessary information to the model:
Start with the default or balanced model. Increase the model or reasoning level when the task actually requires it.
Reserve deeper reasoning for difficult work. Complex analysis, debugging, and multi-step problems may benefit from it. A quick question or simple rewrite usually does not.
Start a new chat when you change topics. Longer conversations can require the model to consider more previous information.
Provide only the material that matters. If you need help with two pages of a 100-page document, supplying those two pages can be more efficient than supplying the entire document.
Be specific in your first prompt. Clearly stating the task, audience, format, and constraints can reduce unnecessary back-and-forth.
Understanding context
Think of an AI model as working at a desk. It can work with the information currently available on that desk, such as its instructions, your current request, relevant conversation history, and information retrieved from available files or tools. The amount of information that can fit on the desk is called the context window, and its capacity is commonly measured in tokens.
This is one reason a focused conversation can work better than an extremely long one. When you start a new conversation, you are giving the model a clean workspace. However, some AI systems may separately provide saved preferences, organizational instructions, or other information to new conversations, so a new chat does not necessarily mean that every AI product forgets everything about you.
Two ways to give an AI tool reusable guidance
Instead of repeatedly explaining the same requirements, reusable instructions can give AI tools guidance about how you want work performed. Two useful approaches are standing instructions and Skills.
| Standing instructions | Skills (SKILL.md) |
|---|---|---|
What they are | Instructions made available broadly or automatically | A reusable bundle of instructions and optional supporting resources |
When they are useful | Rules that should apply consistently | Guidance needed for a particular kind of task |
Examples | Organizational policies, general preferences | Branding, accessibility requirements, document workflows |
Context impact | Broad instructions may consume context even when they are not important to the current task | Detailed Skill resources can be consulted as needed |
What are Skills, and how do they work?
A Skill is a reusable package of instructions that help AI tools perform a particular type of work consistently. Skills use progressive loading; AI tools do not need to read every piece of every installed Skill for every request. Instead, information becomes available as it is needed.
Identify the right Skill. Basic information, including the Skill’s name and description, helps the AI tool determine whether a Skill applies to this request.
Read the instructions. When a Skill applies, the AI tool can use the instructions in its SKILL.md file.
Use supporting resources when needed. A Skill can also contain reference material, scripts, templates, or other resources that are used when a particular step requires them.
This makes Skills especially useful for specialized instructions that would otherwise have to be included in every conversation. For example, an organization might create Skills for applying its writing style, checking accessibility requirements, processing PDFs or following a standardized workflow. Those detailed instructions can remain available without requiring every user to paste them into every prompt.
Additional Resources