Guidelines and Best Practices for Prompt Engineering with Large Language Models

1. LLM Types – The guide distinguishes between a Base LLM (trained on internet data, may give unrelated answers) and an instruction‑tuned LLM (fine‑tuned with prompts and RLHF, can follow user instructions such as “write a journalist‑style introduction about Alan Turing”.)

2. Two Main Prompt Principles – Clarity & specificity : use delimiters (commas, line breaks) and ask for structured output (e.g., JSON). Give the model time to think : break tasks into explicit steps (step 1: summarize, step 2: translate, step 3: extract names, step 4: return JSON) and let the model verify its own answer before responding.

3. Model Limitations – Outputs must be filtered for prohibited content; LLMs can hallucinate. A mitigation strategy is to require the model to cite sources from the provided text.

4. Iterative Development Workflow – For machine‑learning projects: generate an idea → implement code, collect data, train model → inspect outputs, perform error analysis → rethink and repeat. Emphasizes clear prompts, analysis of unexpected results, and continuous prompt refinement.

5. Application Scenarios – Summarization : condense long app reviews into 20‑character snippets for CEOs/CTOs. Reasoning : extract key information from government reports with length constraints. Expansion : turn short texts into longer articles or emails, e.g., brainstorming creative copy.

6. Token Economics – Pricing: $0.002 per 1 000 tokens. A token ≈ 4 characters in English, 2–2.5 characters in Chinese. Roughly 1 000 tokens ≈ 750 words. Typical question consumes 100–200 tokens; continuous conversation increases cost due to history tokens.

7. Online Practice – Demonstrates using the OpenAI API to generate ad copy for a podcast platform. Includes concrete prompt JSON and troubleshooting notes (trim whitespace, reduce batch size, adjust socket timeout). Example API request:

{
    "model": "gpt-3.5-turbo",
    "messages": [
        {
            "role": "user",
            "content": "你现在是一名SEM搜索投放关键词专家，我需要你根据输出给你的文字内容，帮助生产关键词，以使其简练、有重点、让人印象深刻。输出的关键词需要注意：1.能让我快速记住文字内容要点 2.避免对原有文字内容进行改写和添加，只能在原文基础上进行关键词提取 3.结果删除最大、第一、巅峰、优质、经典、精品、全新 4.结果删除免费、畅听、会员、完结、连载、完本、全本 5.结果以标准的Json格式返回 6.结果按照例子分组展示 7.结果字数≤5。例子:0.医林误案 1.大耍儿之两肋插刀 2.投资理财从入门到精通"
        }
    ]
}

Additional role‑based examples show how to set system , user , and assistant messages to maintain context across turns.