This prompt engineering workflow helps you write structured, reliable prompts for large language models like Claude and GPT. It guides you through the four parts of a strong prompt — role, context, task, and output format — and shows when to add few-shot examples or chain-of-thought reasoning for harder tasks. The approach treats prompting like debugging: change one variable at a time and test against several inputs. Use it to turn vague instructions into repeatable prompts. Everything runs in your browser, with no data uploaded.
<context>...</context>). This helps the AI isolate your instructions from your data.Modern Large Language Models (LLMs) are trained on vast amounts of structured documentation. By using a framework like CREATE (Character, Request, Examples, Adjustment, Type, Extras), you provide the "clear mental model" the AI needs to minimize randomness and maximize accuracy.
A strong prompt is structured, not chatty. Most production prompts contain four parts: a role ("You are an expert SQL reviewer"), context (the data, constraints, audience), a precise task (what to do, in what scope), and an explicit output format (JSON schema, bullet list, max length). Vague prompts get vague answers; specifying the shape of the response removes most of the guesswork.
When a task is hard to describe, show examples. One or two input→output pairs (few-shot prompting) usually beats paragraphs of instruction, because the model infers the pattern. For multi-step reasoning, asking the model to "think step by step" before answering improves accuracy on logic and maths.
Iterate like a debugger. Change one variable at a time, test against several inputs (including edge cases and adversarial ones), and pin down what actually moved the result. Constrain outputs you'll parse programmatically to a strict format and validate them. Keep a small library of prompts that work — prompt engineering is mostly disciplined reuse, not magic phrasing.