What is the simple, improved, refined prompt method?

Most people write a prompt once, read the answer, feel mildly let down and move on. The simple, improved, refined method treats the first prompt as a draft. You write it, look at what is missing, add it, and look again. After two or three passes the request says what you actually meant, and the answer follows.

The method comes from A Beginner's Guide to ChatGPT, Robert Scales's book published last month. The book walks through eleven professions, from marketers and accountants to students and programmers, and shows each one's prompt at three levels. For every level it names the problem with the prompt and the outcome you can expect once it is fixed. The pattern is the same every time: simple prompts give generic results, and refined prompts give targeted ones.

The three levels

The improved level answers who and what exactly. The refined level answers how: what the model should do first, what to include, and what the finished answer should look like. Add a role at the top (the subject of the previous article in this series) and you have the book's full method: role, context, task, format, then iterate and verify.

A worked example

Imagine you coordinate volunteers for a small community food bank and need help recruiting.

Simple:

Write a volunteer recruitment post.

The issue: the model has no idea who you are, who you want to reach or where the post will appear. Expect something cheerful and interchangeable.

Improved:

Write a Facebook post recruiting weekend volunteers for our community food bank in a small town. We especially want retirees and high-school students who need community service hours.

The outcome: the post now speaks to two real groups on a specific platform. It will be better, but you still have to decide length, tone and what action people should take.

Refined:

You are a volunteer coordinator with experience in small nonprofits. Write a Facebook post recruiting weekend volunteers for our community food bank in a town of about 4,000 people. We want two groups: retirees, and high-school students who need community service hours. Keep it under 120 words, warm and plain, no exclamation marks. Mention that shifts are three hours on Saturday mornings, that no experience is needed, and end with one clear step: reply to the post or call the number we will add. Then give me two shorter versions to test, one aimed at each group.

The outcome: a post you can publish nearly as is, plus two variants to compare. Notice that none of this is clever wording. Each addition is a fact or a decision you would have had to make anyway.

OpenAI's help article on prompting gives developers the same kind of advice: be specific about context, outcome, length, format and style, and replace fuzzy descriptions with exact ones. Its example swaps "fairly short, a few sentences" for a precise 3-to-5-sentence paragraph.

Refining inside the conversation

You do not have to rewrite the whole prompt each time. The book describes refining as a cycle: read the answer, judge whether it fits, then follow up with whatever was missing. Follow-ups such as these often do the job:

• "Make it shorter and drop the second paragraph."
• "Your list is too general. Give me examples specific to a rural town."
• "That missed the high-school students. Add a line for them."
• "Before you rewrite it, ask me three questions about what's missing."

The last one is worth remembering. When it launched ChatGPT, OpenAI noted that the model usually guesses when a request is unclear instead of asking. Inviting it to ask questions first turns the guessing into a short interview.

How many rounds of refinement are enough?

Enough is when the answer does the job you need, not when it is perfect. The book frames refinement as a cycle of testing, judging the output and adjusting, and it ends when the output meets your goal. For a quick email that may be the improved level and one follow-up. For a report, a lesson plan or a project plan, you may go through the refined level and several follow-ups.

A practical rule of thumb: if two or three rounds of follow-ups have not fixed the problem, stop patching and rethink the prompt. Usually the original request is missing something basic, such as who the answer is for, what the real goal is, or an example of what good looks like. OpenAI's guidance suggests that showing the model an example of the format you want often works better than describing it. Starting a fresh chat with a clean, refined prompt can also help, since a long conversation carries earlier wrong turns along with it.

Finally, keep the prompts that work. The book recommends noting your most effective prompts and what made them succeed. A small personal library of refined prompts means the next similar task starts at level three instead of level one.