What Is Prompt Engineering?

How it works

A large language model combines general knowledge and language patterns learned in training with whatever is in its context for that request: your instructions, any text or files you include, earlier messages, and anything the tool retrieves for you. It won’t reliably know your business, your audience, or your standards unless that information is provided or retrieved. A vague request leaves the model to fill gaps with generic assumptions, which is why the same tool can give one person a weak answer and another a useful one.

Prompt engineering is the habit of closing those gaps on purpose. You decide what the model needs to know, what the output should look like, and what it must avoid, then you test and adjust. Everything the model handles in one request (standing instructions, your prompt, included material, earlier messages, and the reply it writes) has to fit within its context window, the limit on how much text, measured in tokens, it can work with at one time.

Many AI assistants also let you save standing instructions that apply automatically, either across your whole account (often called custom instructions or preferences) or within a single project. They work alongside the system prompt, the behind-the-scenes instructions set by whoever built the tool. Writing either kind well is prompt engineering too, done once and reused.

Practical techniques

Most good prompting comes down to clear writing. These techniques apply across common assistants like ChatGPT, Claude, Gemini, and Microsoft Copilot:

• Give a role, audience, and goal. Tell the model what role to take (for example, “you write emails for a family dental practice”), who the output is for, and what it should achieve, not just the topic.
• Supply source material. Paste the policy, product details, or notes, and ask the model to work only from them. This reduces, but does not eliminate, made-up details.
• Specify format and length. Bullets or paragraphs, approximate length, headings, reading level.
• Show examples. A few samples of good output (often called few-shot prompting) usually convey tone better than a list of adjectives. Say that they show the tone you want, not wording to copy, so the model doesn’t reproduce one too closely.
• Set constraints. Name what to avoid, and say what to do when information is missing, such as asking a question instead of guessing.
• Iterate. Treat the first answer as a draft. Say what to change, and adjust one thing at a time. Because the same prompt can give different results on different runs, try a promising change more than once before you keep it.
• Save templates. Keep prompts that work in a shared place so nobody starts from scratch.

Why it matters for your business

When everyone on a team writes their own prompts, output quality depends on who happened to ask. One person gets a polished draft, another gets something off-brand, and nobody knows why. Shared, tested prompts give the team a consistent starting point. They capture your voice, your rules, and the details that matter, so a new hire can get output much closer to what an experienced colleague gets.

Good prompts also make review easier. When the format and constraints are set, it’s simpler to spot what’s missing or wrong. That review still matters: a well-built prompt makes reliable output more likely, but it doesn’t guarantee accuracy, and the model can still state errors confidently.

Prompts need upkeep. A prompt that works well in one tool may need adjusting in another, and results can shift when a provider updates its model. Revisit your saved templates from time to time and keep the versions that still do the job.

How it compares with fine-tuning and RAG

Prompt engineering changes the input. Fine-tuning changes the model by training it further on your own examples, which adjusts the internal settings the model learned during training (its weights). That difference shapes when each one makes sense.

Prompting is fast, inexpensive, and reversible. You can edit a prompt in seconds and see the effect right away, with no training data or technical setup. Fine-tuning takes prepared examples, technical work, and testing. It’s usually considered for a narrow, repeated task that needs a consistent style or format that prompting alone can’t deliver reliably. Common advice is to get as far as you can with prompting first.

A related approach automates part of the work. With retrieval-augmented generation (RAG), a system finds relevant documents and adds them to the prompt for you, so nobody has to paste source material in by hand. It still depends on good instructions: RAG supplies the context, and prompt engineering shapes what the model does with it.

Further reading

• Claude 101 WorkshopUseful if your group wants hands-on, beginner-friendly practice setting up Claude’s standing instructions, projects, and other personalization features.
• AI Training & SpeakingUseful if you would rather build AI skills as a team through training or a custom workshop than learn on your own.