LLM Fundamentals¶

Seven concepts that explain what an LLM is actually doing between your prompt and its answer — what it is, what it costs, how to control its randomness, and how to instruct it — written for application developers, not model trainers.

flowchart LR A["How LLMs Work next-token, parameters"] --> B["Tokens and Context counting and paying"] A --> C["Temperature and Sampling the creativity dial"] A --> D["Reasoning Models thinking you pay for"] B --> E["Choosing and Tuning families, bake-offs, fine-tuning"] C --> E D --> E F["Embeddings and Vectors meaning as numbers"] --> E E --> G["Prompting and Instructions prompts, AGENTS.md, skills"]

Subtopics¶

#	Subtopic	What you'll learn
01	How LLMs Work	The next-token loop, what "7B" means, why the model never learns from your chat, and why hallucination is structural.
02	Tokens and Context	Tokens vs. words, computing what a request costs, what fills the context window, and why effective context is smaller than advertised.
03	Embeddings and Vectors	How text becomes a list of numbers, cosine similarity, what you can build with it, and when keyword search is better.
04	Temperature and Sampling	What temperature actually changes, why it makes output creative, and when creativity actively breaks your app.
05	Reasoning Models	How "thinking" works, why you pay for tokens you never see, and when reasoning is waste instead of accuracy.
06	Choosing and Tuning	Claude, GPT/Codex, Gemini, GLM, and open-weight families by design intent — plus the prompt → RAG → fine-tune ladder.
07	Prompting and Instructions	The five parts of a good prompt, and the principles behind AGENTS.md, CLAUDE.md, SKILL.md, and rule files.

How to use this section¶

Each subtopic has four levels — junior → middle → senior → professional — retargeted to the application developer: junior means you can use the concept correctly in one small app, middle means you can choose between options for a real product constraint, senior means you can diagnose and design under production constraints, professional means you can set standards for a team. Start at your level. Topics 01 and 02 are the foundation — everything else assumes you know what a token is and what the model is doing. Topics 03–05 can be read in any order after that. Topic 06 pulls the earlier concepts into model-selection decisions, and Topic 07 is the applied discipline of steering the model with written instructions.

For what to put inside the context window at production scale (retrieval pipelines, RAG decisions, compaction), continue to Context Management. For measuring whether any of this actually works, see Agent Evaluation.

Practice rule¶

Before changing a prompt, a sampling parameter, or a model, name the specific failure you're fixing and the number that will tell you it worked. A change you can't measure is a guess wearing an engineering costume.

Part of the AI Agent domain.