Glossary and references

Glossary

Term	Meaning
Prompt	Input that helps determine the model's next response or action; it may include more than plain text.
Prompt Engineering	Designing and evaluating instructions, context, examples, tools, and output contracts for a target behavior.
Token	A model's text-processing unit; token boundaries do not map one-to-one to words.
Context Window	The maximum context a model can process in one interaction/state window, subject to model and API behavior.
Zero-shot	Asking for a task without in-prompt examples.
Few-shot	Supplying a small number of examples to demonstrate a task or output convention.
Chain-of-Thought (CoT)	Intermediate reasoning traces. Historically, prompting models to emit such traces improved some reasoning tasks; modern reasoning models often reason internally, so production prompts should prefer verifiable summaries/evidence over demanding private reasoning.
Reasoning Model	A model family trained/configured to spend additional inference work on complex tasks, often with model-specific reasoning controls.
Self-consistency	Sampling multiple candidate reasoning/answer paths and aggregating them.
ReAct	A research/agent pattern combining model decisions, actions/tool calls, and observations.
Tree of Thoughts	A search/orchestration approach that explores and evaluates multiple candidate reasoning branches.
Structured Outputs	API-level schema-constrained model output, typically using JSON Schema on supported models.
Function/Tool Calling	Allowing a model to select and provide arguments for external functions/tools that the application executes.
MCP	Model Context Protocol, used by compatible systems to expose external tools/resources through standardized servers/connectors.
RAG	Retrieval-Augmented Generation: retrieve relevant external content and provide it to the model before generation.
Prompt Injection	Malicious or unintended instructions that alter model behavior; can be direct or embedded in external content.
System/Developer Instruction	Higher-priority application instructions in APIs that support message-role hierarchies. Not a secure secret store.
Hallucination	Model output that is unsupported, fabricated, or incorrect while being presented plausibly.
Eval	A repeatable test used to measure model/prompt behavior against defined criteria.
Fine-tuning	Updating a model's learned behavior using training data; distinct from supplying examples in a single prompt.
AI Agent	A system in which a model can make decisions across multiple steps and use tools or external systems under an orchestration and permission model.

Key corrections from the source booklet

The supplied v4 booklet is a strong introductory base, but this documentation makes these changes:

• standardizes the term Prompt Engineering instead of the reversed Engineering Prompt wording;
• treats Chain-of-Thought prompting primarily as historical/research context and avoids requiring hidden reasoning from modern reasoning models;
• treats ReAct and Tree of Thoughts as orchestration/search patterns rather than one-line prompt tricks;
• replaces fictional-expertise role prompts with explicit review criteria and perspective framing;
• updates OpenAI examples around the Responses API, Structured Outputs, reasoning.effort, and max_output_tokens;
• avoids universal temperature ranges and other model-agnostic parameter recipes;
• distinguishes ChatGPT product memory from API conversation state;
• states explicitly that the system prompt is not a secret or authorization boundary;
• strengthens RAG guidance around access control, citation validation, and indirect prompt injection;
• makes evals and regression tests part of the prompt lifecycle rather than an optional final review.

Research foundations

The historical techniques described in this guide originate from widely cited research including:

• Brown et al. (2020), Language Models are Few-Shot Learners.
• Wei et al. (2022), Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.
• Kojima et al. (2022), Large Language Models are Zero-Shot Reasoners.
• Wang et al. (2022), Self-Consistency Improves Chain of Thought Reasoning in Language Models.
• Yao et al. (2023), ReAct: Synergizing Reasoning and Acting in Language Models.
• Yao et al. (2023), Tree of Thoughts: Deliberate Problem Solving with Large Language Models.
• Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.

These papers establish techniques in particular model/evaluation settings. They should not be interpreted as guarantees that the same technique improves every current model.

Current operational references

For implementation guidance, prefer current official documentation over static tutorial parameter tables:

• OpenAI Help Center — Prompt engineering best practices for ChatGPT.
• OpenAI Help Center — Best practices for prompt engineering with the OpenAI API.
• OpenAI API Reference — Responses API message roles, reasoning controls, tools, and Structured Outputs.
• OpenAI API documentation — model-version compatibility and eval guidance.
• OpenAI Help Center — current ChatGPT Memory documentation.
• OWASP GenAI Security Project — LLM01:2025 Prompt Injection.
• OWASP GenAI Security Project — LLM07:2025 System Prompt Leakage.
• OWASP GenAI Security Project — current RAG/vector/embedding risk guidance.

Review date

This documentation was technically reviewed against current public references on 2026-08-10. Provider behavior changes quickly; re-check model-specific API documentation before copying parameter names, supported values, or lifecycle dates into production code.