Prompt engineering
Prompt engineering is the deliberate design, structuring, and testing of instructions and examples supplied to a language model so its output more reliably meets a task's requirements. It works at inference time: a prompt can establish a role, specify the task, provide examples, set constraints, and define an output format, but it does not modify the model's parameters. Because model outputs are probabilistic and model-dependent, effective prompting is an iterative engineering activity rather than a one-time wording trick.
Origin and context
Its technical lineage predates consumer chat assistants. A 2021 survey organized prompt-based learning as a paradigm in which inputs are transformed into textual prompts so pretrained language models can perform tasks with few or no labeled examples. By 2023, research on conversational LLMs described reusable prompt patterns for controlling interactions and outputs. The reviewed evidence does not establish a single inventor of prompt engineering; the practice developed across the NLP research and developer communities.
Why it matters
Prompt engineering gives teams a relatively fast way to adapt a general-purpose model to a task without retraining it. It turns desired behavior into reviewable artifacts: instructions, examples, constraints, expected formats, and evaluation cases. In production, the value comes from repeatability rather than clever phrasing. Prompts can be versioned in code, tested against representative inputs, and reevaluated when a model snapshot changes. That makes prompting part of a broader quality loop involving model choice, tests, monitoring, and controlled rollout.
Example
For a support-ticket classifier, a weak prompt might only ask the model to assign a category. A stronger engineered prompt defines the allowed labels, explains ambiguous boundaries, separates the ticket text from instructions, provides a few representative examples, and requires a machine-readable output shape. The team then evaluates the prompt on a fixed test set and records failures before deployment. If the model or prompt changes, the same tests are rerun. This workflow treats the prompt as a testable interface, not as an incantation that guarantees correctness.
How it differs
Context Engineering
Prompt engineering focuses on writing and organizing the instructions, examples, and output constraints presented to a model. Context engineering is broader: it curates the complete token state available at inference time, which may also include retrieved documents, tool definitions, memory, and message history. The concepts therefore overlap, but one does not simply replace the other. A single-turn task may be primarily a prompting problem; a multi-turn agent usually requires context engineering while still relying on well-designed prompts.
Maturity and evidence
Maturity is rated 3: established but still evolving. Prompting has systematic research surveys, reusable pattern catalogs, and current guidance from multiple major model ecosystems, which supports durable use beyond a single vendor. However, effective techniques vary by model family and snapshot, and providers still recommend empirical evaluation when prompts or models change. The practice is therefore mature enough for a stable glossary entry, but its techniques should not be treated as fixed across models.
Limits and open questions
Prompt engineering cannot guarantee factual accuracy, safety, or stable behavior, and it does not update model parameters. Long or overly specific prompts can become brittle across tasks or model revisions. In agentic systems, prompt quality is only one component alongside tools, retrieved data, memory, and conversation state. Teams should use evaluations to decide whether a failure is best addressed by changing the prompt, the surrounding context, the model, or another part of the system.
Related terms
References
- ChatGPT Enterprise: Practical prompt engineering for everyday workOpenAI · 2026-04-27 · class A
- LLMs: Fine-tuning, distillation, and prompt engineeringGoogle for Developers · 2025-12-03 · class A
- Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language ProcessingarXiv · 2021-07-28 · class B
- A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPTarXiv · 2023-02-21 · class A
- Effective context engineering for AI agentsAnthropic · 2025-09-29 · class A
Last updated: 2026-08-27
This term is also covered in the Skills Atlas as prompt engineering skill.
This term is also covered in the Skills Atlas as system prompt design skill.
This term is also covered in the Skills Atlas as in context learning skill.