What is prompt engineering?
Prompt engineering is the practice of designing and refining the text you give to an LLM to reliably get the output you want. It is part craft, part science — you are not writing code, but you are writing instructions that a language model executes.
A small change in wording can mean the difference between a generic answer and a precise, useful one. Prompt engineering gives you a systematic toolkit so you are not guessing.
Why it matters
Without it
Vague prompts produce vague results. You fine-tune or accept hallucinations as a cost of doing business. Every LLM integration feels brittle.
With it
You get consistent, structured, accurate outputs. You spend less time post-processing. Your apps work reliably without needing to fine-tune expensive models.
Good prompts can close most of the gap between a general-purpose model and a fine-tuned one — at a fraction of the cost and effort.
How LLMs process prompts
Understanding what happens under the hood helps you write better prompts. When you send a prompt, the model does not "read" it the way a human does — it converts every word into tokens and processes them through layers of attention.
Tokens, not words
Text is split into tokens — roughly 3–4 characters each. "Unbelievable" is 3 tokens. This matters for pricing (billed per token) and context limits (measured in tokens, not words).
The context window
The model can only "see" what fits in its context window — typically 8k to 200k tokens depending on the model. Everything outside the window is invisible. Your prompt, conversation history, retrieved documents, and output all compete for this space.
Order matters
LLMs are sensitive to the order of information. Instructions placed at the start (system prompt) and end (just before the user turn) tend to have the strongest influence. Content buried in the middle of a long context is more likely to be "forgotten."
Probabilistic, not deterministic
The model predicts the most likely next token at each step. At temperature=0 it is close to deterministic; at higher temperatures it samples more freely. This is why identical prompts can produce different outputs — and why you must test for consistency, not just quality.
Prerequisites
- Basic familiarity with what an LLM is (ChatGPT, Claude, Gemini, etc.)
- No coding required for most pages — concepts are demonstrated with prompts and examples
- Pages 8–10 include Python code snippets using the OpenAI SDK
Series overview
Introduction ← You are here
What prompt engineering is, how LLMs process prompts, prerequisites, and series overview.
Prompt Anatomy
System/user/assistant roles, tokens, temperature, top-p, and context window limits.
Zero-shot & Few-shot
Zero-shot prompting, few-shot examples, how many to include, and formatting them well.
Chain-of-Thought
When CoT helps vs hurts, hidden vs explicit reasoning, and programmatic alternatives.
Role & Persona Prompting
Personas, expert framing, tone control, and audience targeting.
Structured Output Prompting
JSON/markdown/tables, schema-in-prompt, and combining with Pydantic validation.
Prompt Patterns Library
Reusable patterns: Instruction+Context+Format, ReAct, Critic-Refine, Planner→Executor.
Tool Use & Function Calling
Function calling concepts, JSON schema → tool calls, tools vs pure prompting.
Prompt Security & Guardrails
Injection attacks, jailbreaks, data leakage, input sanitization, output validation.
Prompt Testing & Engineering
Test cases, golden datasets, regression testing, accuracy/latency/cost metrics.
Notes
Token counts differ by provider
OpenAI, Anthropic, and Google each use a different tokenizer. A 1,000-word document that fits in 750 GPT-4o tokens may use 850 Claude tokens. Count tokens with the provider's own library before estimating costs or context headroom.
System prompts are re-sent on every API call
The model has no persistent memory between calls. Your system prompt is included in full on every request. In agentic loops with long conversation histories, system prompt size compounds the per-call cost significantly — keep it lean.
Quality gaps between frontier and smaller models are large
Techniques that work reliably on GPT-4o or Claude 3.5 — zero-shot instruction following, structured output, role framing — often fail on 7B open-source models or require few-shot examples. Always test your prompts on the actual target model before deploying.
The field moves fast — check if a newer model handles it natively
Many techniques that required careful prompt engineering in 2023–2024 now happen automatically in extended thinking modes or via native structured output APIs. Before spending time engineering around a problem, check whether the latest model version solves it natively at the API level.
Prompt Engineering FAQ
What is prompt engineering?
Prompt engineering is the practice of designing and refining the text you give to an LLM to reliably get the output you want. A small change in wording can mean the difference between a generic answer and a precise, useful one.
Why does prompt engineering matter?
Good prompts can close most of the gap between a general-purpose model and a fine-tuned one — at a fraction of the cost. Without prompt engineering, vague prompts produce vague results and every LLM integration feels brittle.
Do I need coding skills to do prompt engineering?
No coding is required for most prompt engineering concepts — they are demonstrated with prompts and examples. Only the later pages in this series (covering structured output and tool use) include Python code snippets.
How do LLMs actually process prompts?
LLMs convert text into tokens (roughly 3–4 characters each) and process them through layers of attention. Order matters: instructions at the start and end of the context tend to have the strongest influence. The model predicts the next token probabilistically, not deterministically.
What is a context window in an LLM?
The context window is the maximum amount of text (measured in tokens) the model can see at once — typically 8k to 200k tokens depending on the model. Your prompt, conversation history, and retrieved documents all compete for this space.
Quick summary
- Prompt engineering = designing instructions that get reliable, high-quality LLM outputs
- LLMs work with tokens, not words — order and position in the context window matter
- Good prompts close most of the gap with fine-tuning at a fraction of the cost
- 10 pages: from basics to patterns, function calling, security, and testing