After you have learned the basic role, task, format, and constraint structure, these 10 advanced techniques can give you another level of prompt control. Every one has a concrete example you can copy and try directly.
First, a mindset note: prompt engineering has no “best template,” only “whatever works best for this specific task.” The same task type may need three or four prompt versions, because different models, situations, and even moods produce different results.
1. Assign Multiple Roles
Do not give the AI just one role. Stack two or three identities.
Single role: “You are an SEO consultant.” The AI will answer from an SEO angle, but it may ignore copy quality.
Stacked roles:
你同時具備以下三種身份：
1. 資深 SEO 顧問，熟悉 Google 2026 年的排名因素
2. 台灣科技部落客，擅長用白話解釋技術
3. 嚴格的編輯，會砍掉所有不必要的形容詞
用這三個角色的綜合判斷來完成以下任務。This makes the AI consider SEO, readability, and tighter writing at the same time. For blog article prompts, a three-layer role setup works especially well.
2. Send Instructions in Chunks (Chunking)
Do not dump a long task into the chat all at once. Split it into two or three parts, and only continue after confirming the AI understands each part.
Send the background and rules first:
想寫一篇關於 RAG 的教學文章。先不要寫，會分段提供資訊。
以下是這篇文章的目標讀者和風格要求：
[貼上風格規則]
請確認你理解了，用一句話摘要你的理解。After the AI replies with confirmation, send the outline and materials in the second step. Only in the third step do you ask it to start writing.
This move works especially well with ChatGPT. Its attention can drift inside a long prompt, and chunking keeps it focused at each step.
3. Give Examples to Calibrate Output (Calibration Shot)
Before the real task, give the AI a small test.
在開始正式任務之前，請先用以下範例校準你的風格：
範例輸入：「什麼是 API？」
範例輸出：「API 就是兩個軟體之間的溝通橋梁。你用手機查天氣，手機 App 就是透過氣象局的 API 拿到資料的。不需要懂背後的技術，只要知道它是一個讓不同系統互相交換資訊的管道。」
請用這個風格回答接下來的問題。One calibration example beats a 500-word style description. AI picks up style signals from examples much more precisely than from abstract wording.
4. Self-Check Before Finishing
Ask the AI to review its own work after writing.
完成以上任務後，請自行檢查：
1. 有沒有超過字數限制？
2. 有沒有使用以下禁用詞：[列出]
3. 每個論點是否都有具體例子支撐？
如果有問題，直接修正後輸出最終版。不需要列出修改過程。Experience note: Claude is relatively strong at self-checking. Adding a self-review instruction at the end of a prompt is like putting one more quality gate in the workflow.
5. Provide Negative Examples (Negative Few-Shot)
Tell the AI “do not write like this” by pasting a negative example directly.
以下是不好的輸出範例，請避免這種風格：
不好的範例：「在 AI 快速演進的時代，選擇合適的模型至關重要。Claude 和 ChatGPT 各有優劣，前者擅長長文撰寫，後者則在創意發想方面表現突出。」
問題：開頭太八股、語氣像報告、沒有個人觀點、用了「至關重要」「各有優劣」。
請用朋友聊天的語氣重新表達同樣的資訊。Positive examples teach the AI what you want. Negative examples teach it what to avoid. The two work best together.
6. Style Guidance
Chat interfaces do not let you directly adjust model temperature, but you can influence the AI’s “creativity level” with wording.
When you need precision: “Answer strictly based on the provided material. Do not speculate. Mark uncertain parts as ‘to verify.’”
When you need creativity: “Think boldly. List 10 unusual angles. Do not worry about feasibility. The more unexpected, the better.”
The same model can produce very different output under these two instructions. When planning SEO content topics, use “creative mode.” When writing the main draft, switch to “precise mode.”
7. Structured Input
Write the prompt itself in a structured format so the AI can parse it more easily.
# 角色
台灣的 AI 教育工作者，政大資工碩士畢業
# 任務
寫一篇 1500 字的教學文章，主題：RAG 入門
# 讀者
- 非技術背景的上班族
- 聽過 AI 但沒用過 RAG
- 年齡 25-40 歲
# 格式
- Markdown
- 用 H2 和 H3 分章節
- 每章節 200-400 字
# 限制
- 繁體中文
- 不用「淺顯易懂」「華麗轉身」這類空話
- 不超過 1500 字Using Markdown headings to separate sections helps AI understand each block’s intent more accurately. The parsing error rate drops a lot compared to dumping everything into one wall of text.
8. Iterative Refinement
If you are unhappy with the first output, do not start over. Give specific revision instructions.
這個版本有三個地方要改：
1. 第二段的開頭太像報告，改成聊天語氣
2. 第四段的例子不夠具體，請用台灣企業的案例替換
3. 全文太均勻，把第三段縮短到兩句就好
其他部分維持不動。A lot of people delete the whole prompt and rewrite it when the first version disappoints them. Refining what already exists is faster than starting again, and the AI still has the earlier conversation context, so the revisions land closer to what you want.
9. Lock the Output Format
Give the output format skeleton at the beginning of the prompt.
請用以下格式輸出：
## [章節標題]
[1-2 句核心概念]
[具體範例，用引用區塊格式]
[小企鵝的觀察：1-2 句個人心得]
---
每個章節都按照這個格式。一共 5 個章節。The benefit of format anchoring is that AI repeats the same structure very consistently. It is especially useful for series articles, comparison tables, FAQs, and other repetitive outputs.
10. Ask AI to Generate the Prompt for You
The most advanced move: ask AI to help write the prompt.
想完成以下任務：[描述任務]
請寫一個最適合 Claude 的 prompt，包含角色設定、任務描述、格式要求和限制條件。
寫完之後解釋你為什麼這樣設計每個部分。AI-generated prompts are usually more structurally complete and detailed than prompts people write from scratch. Before the real work starts, you can still review each detail and adjust it before moving forward.
Prompt Methodology: Three Principles
After using prompts for a long time, I keep coming back to three principles.
Run first, then revise. Do not chase perfection on the first try. Write a prompt that is 70% done, run it once, check the result, and revise the parts you dislike. That is more efficient than spending 30 minutes writing a “perfect prompt.”
Prompts are alive, so frequently used prompts or prompt libraries should be updated regularly. Models change, and the same prompt can behave differently across Claude Opus versions. I recommend reviewing a prompt library every month or two: remove outdated prompts and update the ones whose quality has slipped.
Log the screwups. Whenever AI output is far from what you expected, note which part went wrong and how you fixed it. The pitfalls you have already hit become your most valuable prompting experience.
Common Mistakes
Three common mistakes to close with.
Stacking too many roles. Three role layers can work. Eight layers become chaos. The AI will not know whose priority to follow, and the output goes all over the place. Three layers is the ceiling.
Contradictory constraints. If you put “be creative” and “strictly follow the example format” together, the AI gets stuck. For creative tasks, loosen format constraints. For precise tasks, tighten the creative space. Do not demand both at the same time.
Ignoring model differences. A prompt that works beautifully in Claude may drop to about 60% as good in ChatGPT. Every model has its own temperament, so it is worth maintaining a separate version for the models you use often. I cover those differences in Claude vs ChatGPT and Gemini vs ChatGPT.
You do not need to learn all 10 techniques in one sitting. Pick two or three that match your work most closely and start using them. Once those become reflexes, come back for the rest. You can also start from the basic structure in the complete prompt guide.
Penchan’s Experience
I have a fixed prompt library, mainly for image generation, built from long-running templates I use often. Whenever a prompt reliably gives me usable output, I file it by category and reuse it next time with a few tweaks.
From what I have seen in practice, the most common combo is “role stacking + structured input + self-check.” For blog articles, I stack three roles: domain-specific consultant, plain-language translator, and strict editor. I use Markdown blocks to make the prompt clear, then add a final paragraph asking the AI to check banned words and word count by itself. That combo has the most consistent hit rate per article for me.
For model differences, Claude is best at following long rules and banned-word lists. I can list more than a dozen banned terms and it will avoid them. ChatGPT starts selectively ignoring constraints when the list is too long, so I rely on few-shot examples to anchor its style. Gemini fits long-context tasks with lots of data, but I have to be careful with Chinese-language feature limits; sometimes switching the prompt to English is what triggers certain features.
I review the prompt library every month or two. After a model version changes, the same prompt can behave differently. If I do not set aside time for maintenance, the library turns into a pile of outdated templates.
Further Reading
FAQ
Q: What is the difference between advanced prompt tips and the basics?
The basics combine four elements: role, task, format, and constraints. Advanced prompting adds strategy on top of that foundation, such as calibrating style with examples, breaking a task into stages, or asking AI to review its own output. The difference is how precisely you can control the result.
Q: Do these tips work with every AI model?
The general direction works across models, but the effect varies. Claude responds best to long rules and constraints, ChatGPT is most sensitive to learning from examples, and Gemini fits large datasets paired with search. Each technique in the article notes which model it suits best.
Q: How long does it take to learn these techniques?
You can understand each technique in about five minutes after reading the example. To use them smoothly, you need to edit the examples yourself, run them a few times, and learn how the model reacts. After a week or two, the ones you use most often usually become habit.
Q: Will a long prompt exceed the token limit?
Mainstream models in 2026 have very large context windows. Claude has 200K tokens, and Gemini has 1 million tokens, so long prompts are not technically a problem. The real question is whether the length is useful, not whether it is long.
Q: What is the best way to practice prompting?
The most effective practice is to take one real task from your daily work, write a prompt, run it once, check what you dislike about the result, revise the prompt, and run it again. Doing that loop ten times teaches faster than reading tutorials.
— Penchan