Prompt Engineering
Prompt Engineering
What is prompt engineering?
- Designing instructions that get the results you want
- Key elements: role, context, examples, and format
- The basics of working with AI
Once prompt engineering takes hold, the team’s conversation changes. Evaluations that used to end with “that AI isn’t great” become “show me the screen. What instruction did you ask it with?” The real value of this skill is shifting responsibility for results from the model to the instructions. A well-refined prompt becomes a document the team copies and reuses, not one person’s knack, so anyone who runs it gets results of similar quality.
In practice, it helps to split instructions into three layers. Things that rarely change, such as company background, tone, and banned words, go in the system prompt. The purpose and audience of the current task go in the user prompt. For tasks where the judgment criteria are ambiguous, add two to five examples (few-shot). For calculations or multi-step reasoning, asking the model to show its process, as in “work through it step by step, then output only the conclusion,” noticeably reduces wrong answers. Still, you need to separate what a prompt can solve from what it cannot. If answers must be grounded in internal documents, RAG is the right tool; if you need to lock in tone and format at scale, fine-tuning fits better. Rewriting instructions alone cannot create facts that were never there.
The most common misconception in practice is that you only need to collect “prompts that work.” People paste in lines copied from online communities and, if the result is disappointing, go looking for another, yet often not a single line describes our own customers or constraints. Simply making prompts longer is also a trap: attach twenty pages of reference material and the key instruction gets buried, so the AI summarizes the wrong items. Existing prompts can break when the model is updated, which is easy to overlook, so it is safer to rerun 5 to 10 representative prompts every quarter. Finally, the habit of pasting customer lists or unpublished results as they are leads not to a performance problem but to a security incident.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Seolyeong ParkGenerative AI and AX training specialist with 17 years of hands-on experience and 200+ lectures, helping non-technical staff put AI to work right awayA practical skill that emerged in the early 2020s, as large language models (such as GPT-3) went mainstream and people realized that how you give instructions determines performance. It took shape quickly within user communities rather than through any single founder.
Consider a hypothetical case. The marketing team at Company A, an industrial supplies distributor with KRW 62 billion (≈ US$46,386,400) in annual revenue and 210 employees, had three people producing 30 product detail page descriptions a month. Four months after adopting generative AI, time per piece was still 4 hours, and the team lead sent each piece back an average of 2.4 times.
- When Ms. Kim, the marketing team lead, collected the instructions her team had actually typed, nine out of ten were a single line: “Write copy for this product detail page.” “Anyone would end up with something bland that way,” was her first reaction.
- So they turned the company introduction they kept repeating, the interests of equipment purchasing managers, and seven banned internal phrases such as “industry No. 1” into a fixed block at the very start of every prompt.
- We also specified the output format. Once we wrote “three headline options of 15 characters or fewer, one each from the angles of delivery time, unit price, and after-sales service, output as a table,” the first draft needed far less editing.
- We fixed the tone by attaching last year’s three highest click-rate phrases as examples. Ms. Lee, a first-year hire, was amazed: “It sounds exactly like how I normally write.”
- The four finished prompts were posted to the internal wiki with version numbers (v1, v2), and new hires were asked to run them themselves as a day-one onboarding assignment.
- Six weeks later, time per piece fell from 4 hours to 1 hour 20 minutes, revision rounds dropped from 2.4 to 0.8, and monthly output rose from 30 pieces to 48.