The presentation gives you the map. The guide below explains what each part means when you are actually working.
1. Tokens: the units the model reads
A token is a chunk of text, often about three quarters of an English word. You do not normally see tokens, but they determine how much fits in a conversation and, for API users, how much a request costs. A useful rough rule is 1,000 tokens ≈ 750 words.
2. The context window: the model’s desk
The context window holds the company’s hidden instructions, uploaded files, the conversation and the answer being written. Long chats can become less reliable because the important material competes with everything else. Upload the pages that matter, repeat crucial requirements, and begin a fresh chat when the conversation has wandered.
3. Three different limits
Context limit is the size of the desk. Usage limit is how many messages your plan allows in a period. Rate limit is mainly an API restriction on requests per minute. They are different problems and need different fixes.
4. Prompting that actually helps
Give the model the material it needs. Show an example of the desired result. State measurable boundaries such as “maximum 150 words.” Put the task after a long document so it is close to the answer. For difficult work, ask what information is missing before the model starts. Most importantly, treat the first response as a draft and refine it.
A reusable prompt shape
TASK: What I want done
CONTEXT: Background and source material
RULES: Tone, length and constraints
FORMAT: The exact shape of the answer
5. Why markdown works well
Markdown is plain text with simple structural signals such as # for headings and - for lists. Models usually read it reliably. PDFs may lose columns or table structure during extraction, while screenshots require visual reading and introduce more chances for error.
6. Saving repeated context
Put recurring information in custom instructions, a project, or an instruction file such as CLAUDE.md or AGENTS.md. RAG—retrieval-augmented generation—handles very large collections by finding a few relevant fragments and placing those in context. The model is only as well-informed as the fragments retrieved.
7. Tools, connectors and MCP
A model produces text. Tools let the surrounding application search, read calendars, inspect files or take other actions. The model requests a tool in a structured form, the application runs it, and the result returns to the model. MCP, or Model Context Protocol, is an open standard for connecting those tools to AI applications.
8. Skills: reusable procedures
Skills contain instructions and sometimes scripts or templates. Only the relevant skill needs to load for a task. Tools give a model new abilities; skills give it repeatable procedures.
9. Agents: models in a loop
An agent chooses an action, uses a tool, examines the result and decides what to do next. Repetition makes small errors accumulate, so agents work best with narrow jobs and objective checks: tests, builds, previews or clear acceptance criteria. Anything an agent reads may contain malicious instructions, so external content must be treated as information—not authority.
10. Match the model to the job
Small models are useful for reformatting and simple summaries. Larger models suit analysis, planning and code. Reasoning models are valuable for multi-step logic or mathematics. Use the lightest model that matches the cost of being wrong.
11. What AI is genuinely bad at
AI can be confidently wrong. It is fragile with exact quotes, citations, statistics, recent events, counting and arithmetic performed only in text. Ask it to search current sources when needed and to run code for calculations. Open the cited sources yourself.
12. Protect what you paste
Do not paste passwords, API keys, client records, medical information or other people’s private data without understanding the product’s data policy. Business plans often have stronger protections, but “business” is not a substitute for checking.
The short version