Prompt Engineering

The problem

You have a ChatGPT or Claude subscription. You typed “help me write a customer email,” got something bland, and concluded the tool is overrated. Someone in the next office is running the same model and producing usable drafts in one pass. The difference is almost never the subscription tier — it is what got typed into the box. Most people brief an AI model worse than they would brief a temp worker on their first morning, then blame the worker.

What it is

Prompt engineering is writing the instructions you give an AI model so it returns what you actually wanted. A “prompt” is the text you type in: a question, a request, some context, or an example of the answer you want. The “engineering” is structuring that text on purpose — picking the words, naming the audience, stating the format, showing an example, listing the constraints — instead of typing the first thing that comes to mind. OpenAI: “Prompt engineering is the process of writing effective instructions for a model, such that it consistently generates content that meets your requirements.” — OpenAI, Prompt engineering, 2026. Anthropic’s mental model has stuck with non-technical readers: “Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result.” — Anthropic, Prompting best practices, 2026. It is not programming and not magic phrasing — it is briefing. LLM stands for large language model: the AI behind these chat tools, trained to predict and generate text.

Origin note: The discipline’s most influential technique came from a Google Brain paper on 28 January 2022 — “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” by Jason Wei and colleagues, showing that asking a model to work through intermediate steps before answering dramatically improved its reasoning. OpenAI’s instruction-tuned models arrived the same month.

How it works

Eight steps, distilling what the major labs actually agree on. No code required.

1. State the goal in one sentence. “Help me with my email” is not a goal. “Rewrite this customer complaint so it acknowledges the issue, takes responsibility, and offers a 20% refund — under 120 words” is. Specify success first.
2. Give the model a role. One sentence of persona changes tone: “You are a senior customer-success rep at a small DTC brand.”
3. Specify the format. Bullet list, table, three paragraphs, under 100 words. Models default to a shape you probably do not want.
4. Show, don’t just tell. Paste one or two short examples of input and the output you want. Few-shot prompting, the most reliable way to lock in style. Anthropic: “A few well-crafted examples (known as few-shot or multishot prompting) improve accuracy and consistency. When adding examples, make them: Relevant, Diverse, Structured.” — Anthropic, Prompting best practices, 2026.
5. Add constraints and context. What must appear (brand name, price, disclaimer); what must not (no medical claims, no competitor mentions).
6. Ask it to think first, then answer. For anything with steps — arithmetic, logic, planning, diagnosis — add “Before answering, think step by step, then give the final answer.” Chain-of-thought, from the 2022 paper.
7. Iterate, one variable at a time. Two or three small rewordings, compared. Change the example, the constraint, or the context — not all three.
8. Pin the version and build a tiny test set. Freeze the model, keep five to ten example inputs, and notice when updates break it. OpenAI: “Pinning your production applications to specific model snapshots… Building tests and evaluation suites that measure prompt behavior so you can monitor performance as you iterate, or when you change and upgrade model versions.” — OpenAI, Prompt engineering, 2026.

The single best sanity check in any of the published guides is Anthropic’s: “Golden rule: Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, Claude will be too.” — Anthropic, Prompting best practices, 2026. If a human cannot execute your brief, a model will not either.

One more thing worth internalising, because it tells you how much instruction to write. OpenAI draws the line by model type: “A reasoning model is like a senior co-worker. You can give them a goal to achieve and trust them to work out the details. A GPT model is like a junior coworker. They’ll perform best with explicit instructions to create a specific output.” — OpenAI, Prompt engineering, 2026. Two sentences, the whole trade-off: with a reasoning model, over-specifying can get in the way; with a standard GPT-style model, vagueness is what kills you. Check which one you are talking to before deciding how much detail to write.

Where it works

• Writing and editing at volume. Product descriptions, marketing emails, customer replies, social captions.
• Pulling structured data out of messy text. Names, dates, line items, sentiment from emails, reviews, and call transcripts. Google’s Gemini docs: “For the given order, return a JSON object that has the fields cheeseburger, hamburger, fries, or drink, with the value being the quantity.” JSON is a standard machine-readable text format; the point is output a spreadsheet or database can ingest directly.
• Multi-step reasoning. Arithmetic, planning, root-cause analysis, debugging. Chain-of-thought earns its keep: “Prompting a 540B-parameter language model with just eight chain of thought exemplars achieves state of the art accuracy on the GSM8K benchmark of math word problems, surpassing even finetuned GPT-3 with a verifier.” — Jason Wei et al., Google Brain, 2022. GSM8K is a grade-school maths word-problem test; eight worked examples in the prompt beat a specially retrained model.
• Role-based expert assistance. A Python helper, a contract reviewer, a marketing strategist. Assigning the role up front reliably raises depth.
• Repeatable business workflows. The same prompt run thousands of times with a pinned model and a small test suite catching drift. Boring, valuable, and where most operators make money.

Where it breaks

• Live facts and anything past the training cutoff. “What’s the share price right now?” is not a prompting problem — no wording fixes it, and the model may invent an answer. Fix: use a tool with web or database access, and check that it retrieved something rather than guessed.
• High-stakes decisions with no human reviewing them. Medical, legal, financial — every major lab warns against it. Fix: keep a named human accountable before output reaches a customer or a regulator.
• Tasks the model is not capable of. Prompting cannot manufacture capability. Lilian Weng: “The benefit of CoT is more pronounced for complicated reasoning tasks, while using large models (e.g. with more than 50B parameters). Simple tasks only benefit slightly from CoT prompting.” — Lilian Weng, Prompt Engineering, 2023. Fix: try a stronger model before rewriting the prompt a ninth time.
• Brittleness under rephrasing. Same intent, different wording, wildly different output. Weng again: “the order of the examples can lead to dramatically different performance, from near random guess to near SoTA.” Fix: freeze the wording of prompts that work and change one element at a time.
• Stated reasoning you cannot trust. A model’s visible “chain of thought” may not reflect how it actually reached the answer. Fix: verify the conclusion against a source, not the explanation.

Weng, who has worked inside both OpenAI and Anthropic, is blunt: “It is an empirical science and the effect of prompt engineering methods can vary a lot among models, thus requiring heavy experimentation and heuristics.” — Lilian Weng, 2023. Treat every “rule,” the eight steps above included, as something to test, not a law of nature.

There is also a cost. Every example and “think step by step” adds text; Google notes adding examples “increases the token count and may increase the latency and cost of the call.” Tokens are the units of text a model bills you for.

What this article does NOT cover

• Fine-tuning, retrieval systems, or anything that changes the model rather than the input.
• Writing code against an API — everything here works in a normal chat window.
• Model-by-model comparisons or which subscription to buy.
• Building autonomous agents that call tools and act without a human approving each step.

Sources

• Prompt engineering — OpenAI platform documentation
• Prompting best practices — Anthropic Claude documentation
• Chain-of-Thought Prompting Elicits Reasoning in Large Language Models — Wei et al., Google Brain, 2022
• Prompt Engineering — Lilian Weng, Lil’Log (2023)
• Prompt Engineering Guide — DAIR.AI
• Prompt design strategies — Google Gemini API documentation