Prompt engineering — the full vocabulary
Chain-of-thought, few-shot, ReAct, DSPy, and every prompting technique. The vocabulary of the craft that turned LLMs into practical tools.
Prompt engineering is the practice of designing inputs that get the best output from a language model. Once dismissed as "not real engineering," it has become essential craft — every production LLM system depends on prompts, and small prompt improvements often outperform expensive fine-tuning.
The 40 terms here are grouped into four sub-topics: fundamental techniques (few-shot, CoT, ReAct), frameworks and tools that systematize prompting (DSPy, TextGrad, structured output libraries), advanced patterns (meta-prompting, self-consistency, tree-of-thoughts), and the anti-patterns and attack vectors you need to know (prompt injection, jailbreak, sycophancy). For depth, see our Master Prompt Engineering path.
The most important terms in Prompt Engineering
Start here if you're new. These entries explain the foundational vocabulary in depth.
Prompt
The complete input given to a language model — instructions, context, examples, and query. Effective prompts shape output quality more than most other levers.
Read the full entry →TechniquesChain-of-Thought (CoT)
Prompting the model to reason step by step before answering. Introduced in 2022; foundation for modern reasoning models.
Read the full entry →TechniquesFew-shot Prompting
Giving the model 1-10 example input-output pairs before your actual query. Often outperforms zero-shot for tasks with specific format requirements.
Read the full entry →StructureSystem Prompt
Instructions that shape model behavior before user messages arrive. Sets persona, constraints, output format, and defaults.
Read the full entry →AdvancedReAct Prompting
Reasoning + Acting — alternating between thinking steps and tool calls. Foundation of most agentic patterns.
Read the full entry →FrameworksDSPy
Stanford's framework that treats prompting as programming. Optimizes prompts automatically rather than hand-crafting them.
Read the full entry →All Prompt Engineering terms, organized
Grouped into sub-topics so you can find neighbors and prerequisites, not just alphabetical entries.
Fundamental techniques
Advanced patterns
Frameworks & tools
Anti-patterns & attacks
Related term categories
These categories connect naturally to Prompt Engineering — many terms cross-reference between them.
Go deeper on Prompt Engineering
📖 Related concept tutorials
Long-form guides that walk through how these concepts actually work.
- Prompt Engineering Guide
The complete techniques walkthrough.
- Understanding Prompt Injection
The #1 agent security issue.
- How Chain-of-Thought Actually Works
Under the hood explanation.
🎯 Related models & tools
Real products and models where you'll encounter these terms.
- AI tools for prompting
LangChain, DSPy, Promptfoo, LangSmith.
- Master Prompt Engineering
8-lesson learning path.
- Prompt vs Fine-tuning
When to reach for each.
Questions about Prompt Engineering
The bar is rising, but the discipline is not disappearing. Newer models need less careful prompting for simple tasks — but production systems, agents, and edge cases still require deliberate prompt design. The nature of prompt engineering shifts from 'clever tricks' to 'systematic design.'
Less so. Reasoning models (o-series, Claude with extended thinking, DeepSeek R1) do CoT internally. You still benefit from explicit CoT for complex tasks the model wasn't trained to reason through by default.
Few-shot puts examples in the prompt at inference time. Fine-tuning updates the model weights. Few-shot is faster to iterate on but uses tokens every request; fine-tuning is a one-time cost. See our comparison.
Two things: read what works (papers, others' prompts, provider cookbooks) and evaluate systematically. Every good prompt engineer maintains an eval set for their common prompts and treats prompt changes like code changes.
No. Modern models have some protections (Claude's XML-tag isolation, model-level defenses) but no full solution exists. Assume any content that came from a user or the internet could contain injection attempts.