A bilingual highlight of Google’s official Prompting Essentials course. This is a set of standardized, accessible-to-everyone prompt frameworks that require no technical background, positioned to “consistently get high-quality results from generative AI.”
When to open this:
- You are new to using AI and want a foolproof “prompt template” skeleton you can apply directly
- Your prompts never get the results you want, and you want to know which direction to adjust them (the 4 iteration methods)
- You have a multi-step large task (writing a marketing plan / building an onboarding course / writing an analysis report) and want to break it into controllable small steps (prompt chaining)
- You are looking for reusable playbooks for specific office scenarios (emails / meeting notes / spreadsheets / data charts / text-to-image / presentations)
- You need a checklist to teach your team “how to use AI responsibly”
Complementarity with newtype · Prompt / Context Engineering in Practice: Lao Huang’s piece presents the contrarian position that “prompt engineering doesn’t need to be specifically studied; its essence is task decomposition + writing briefs” plus product-grade prompt orchestration (Prompt House / Ultra think / structured JSON text-to-image). This article is : Lao Huang's piece takes the counter-consensus position that "prompt engineering does not need to be learned specifically; its essence is task decomposition + Brief", plus product-level Prompt scheduling (Prompt House / Ultra think / structured JSON text-to-image). This piece is Google’s official standard framework and scenario templates, more suitable for zero-beginners to use right away. The conclusions of the two are actually consistent — the core is both “state requirements clearly as if onboarding a new hire” — and they can be read cross-referenced.
1. Core: The 5-Element TCREI Prompt Framework
Why this is the foundation of the entire course: 90% of cases where “AI answers poorly” are rooted not in the model being stupid, but in the input being too vague. This framework turns “writing a prompt” into a checkable checklist — just as when you ask a new hire who writes well but knows nothing about your company for help, you must clearly explain what to do, what the context is, what references look like, then verify the result and request revisions. The framework itself is not new knowledge; it standardizes “how to onboard a new hire.”
Mnemonic: Thoughtfully Create Really Excellent Inputs = five initials = five elements.
| Element | English | What to write in the prompt | Understand in one sentence |
|---|---|---|---|
| Task | Task | What you want AI to do + persona + output format | “You are X, use Y format, help me do Z” |
| Context | Context | Goal, reason, what has already been tried, who the audience is | Provide enough detail; the difference is all in the details |
| References | References | Examples / model essays / style samples (few-shot) | “Do it like this one” |
| Evaluate | Evaluate | Acceptance check: Is it accurate? Relevant? Are there hallucinations? | Humans act as gatekeepers; do not trust blindly |
| Iterate | Iterate | If unsatisfied, add information / rephrase and try again | Always Be Iterating(ABI) |
Key reminder:Order does not matter; clear content does. The course repeats one mantra — “the more careful, specific, and clear your input, the better the output usually is.”
Reusable skeleton (apply directly)
任务 Task:
你是一名【角色,如:专攻意大利电影的影评人 / 有 15 年经验的市场主管】。
帮我【动词开头的任务,如:列出 / 起草 / 生成…】。
输出用【格式:表格 / 项目符号 / 三段短文…】。
背景 Context:
【为什么做这件事 + 受众是谁 + 已尝试过什么 + 任何关键约束】
参考 References:
请参考以下示例的【语气/风格/长度】:
<example01> 【范文1】 </example01>
<example02> 【范文2】 </example02>
（用过渡句"参考以下材料"或 XML 标签包裹,让 AI 知道参考从哪开始）
评估 Evaluate / 迭代 Iterate:
拿到结果后自己检查准确性,不对就补细节再来一次。
Two bonus points for writing tasks(emphasized by the course):
- persona — have AI answer from the perspective of a certain domain, automatically filtering out irrelevant content (e.g., “as a science expert”).
- format — specify bulleted list / table in advance to save rework.
4 ways to delimit reference materials(from light to heavy): transition sentence (“Refer to these materials”) → sectioned headings → XML tags (<example01>…</example01>) → Markdown markers (use ** _ to tell AI to preserve the emphasis structure of bold and italics).
Analogy for a 5-year-old (verbatim from the course): Task = “Mom, I want strawberry cake”; Context = “Because it’s my birthday today”; References = “Like the pink one last time”; Evaluate = “Check if it’s what I want”; Iterate = “More strawberries.” First state what you want → then explain → give examples → check if it’s right → revise if wrong.
2. 4 Iteration Methods: Where to Adjust When Your Prompt Doesn’t Work
Why a separate section: The biggest roadblock for beginners is not “writing the first version,” but “not knowing what to do after the first version fails,” so they either start over from scratch or give up. The ABI mindset given in the course is small-step adjustments, and it provides 4 clear modification directions, turning “blind tweaking” into “ordering from a menu.”
| Method | English | When to use | How to adjust |
|---|---|---|---|
| 1 Return to the framework | Return to the framework | The result is too generic, sounds like nonsense | Check if Task/Context/References are missing items; add persona + audience + format |
| 2 Split long prompts | Break down long prompts | One prompt crams in 3 things: “summarize + make image + extract” | Split into Step1 summarize → Step2 make image → Step3 extract; do only one thing at a time |
| 3 Switch to a similar task | Analogous task | AI consistently writes poorly for a certain phrasing | Change the angle, e.g., “write a marketing plan” → “write a story: how the product fits into users’ lives” |
| 4 Add constraints | Constraints | The result is too divergent, not focused | Add limits, e.g., “only artists from a certain region” “only published in the last 5 years” |
Example comparison (return to framework):
- ❌ Before optimization:
Give me five blog post ideas.(Give me five blog ideas) - ✅ After optimization:
You are an expert on sports nutrition. Provide five blog post headlines summarizing the biggest industry trends, for an audience of physical therapists working with professional basketball players.(added persona + audience + specific direction)
Advanced: Chain of Density summarization
Why this trick is valuable: Asking AI to “directly give the shortest summary” in one sentence is dangerous — you can’t see what it deleted or whether it made things up. Chain of Density is gradual compression, where you can check at each level whether key information is still present.
- How to do it: First ask for a detailed summary → “compress into two or three core trends” → “compress into one sentence highlighting the main trend.”
- Or ask for 3 density versions at once:
A detailed summary / A medium-length summary / A one-sentence summary, compare side by side and pick one to use. - Two benefits: ① Stepwise compression lets you detect hallucinations (check facts when compressing from detailed to short); ② See clearly which information is lost during compression, and judge whether important content was dropped.
- The reverse also works: If the answer is too brief, do the reverse
Add more details / Expand the background, sliding freely between “concise” and “detailed.”
3. Prompt Chaining: Break Large Tasks Into a Sequence of Small Steps
Why a separate section: Many projects are too big to finish with one prompt (after finishing a novel you need to make a marketing plan, design a full onboarding course). Prompt chaining is like doing a jigsaw puzzle — instead of dumping all the pieces out and fitting them randomly, you first do the corners, then the edges, then fill in the middle, building step by step. This is the core approach for handling “multi-step complex tasks.”
Definition: Use a sequence of interconnected prompts to guide AI to complete complex tasks step by step.Use the output of each step as the context for the next step’s input.
Key difference from “iteration”(explicitly distinguished in the course):
- Iteration = repeatedly revising the same prompt to get a better single-step result.
- Prompt chaining = each new prompt builds on the previous step’s output, strung into a continuous chain.
Reusable skeleton: Marketing plan chain (original course example)
Scenario: The novel is finished, and you need to promote it. Leveraging the long context window (the amount of information the model can remember at once), the model remembers the entire manuscript, and you only need to issue new tasks afterward.
Step 1（宣传语）：为这本小说手稿生成 3 个一句话总结，语气贴合原稿但更抓人。
Step 2（优化）：把前 3 个组合成一句宣传语，突出情节反转与悬疑，简洁且让人想读下去。
Step 3（封底简介）：基于一句话总结，扩写成 5 句话的封底 blurb。
Step 4（推广计划）：生成 6 周书籍巡回推广计划，含去哪些地点、每站用什么渠道。
Another skeleton: Onboarding course chain — generate course outline → expand the 2nd key point into a 100-word paragraph → create a 3-question quiz based on it → generate a congratulatory image for answering correctly. The same chain spans both text and image modalities.
Two advanced variants (from Key takeaways)
- Chain-of-thought: Have AI explain how it arrived at its conclusion, making reasoning more transparent and verifiable.
- Tree-of-thought: Have AI propose multiple different options at the same time and compare them(e.g., “Imagine 3 designers each making a proposal, each writing one step of their thinking then sharing; whoever finds they are wrong drops out”), suitable for creative scenarios that require divergence + selecting the best option.
4. High-Frequency Office Scenario Playbooks (Each With a Reusable Skeleton)
Why this section is the most practical: Frameworks are the foundation, but what PMs actually do every day is concrete work. The 6 + N scenarios below each come with a prompt skeleton you can directly adapt by changing the fields.All playbooks are fundamentally based on the TCREI in §1, just with fields swapped per scenario.
4.1 Brainstorm
Why: When you need creativity, new ideas, or are stuck, treat AI as a conversational creative partner, adding information and guiding as you go, rather than expecting a perfect answer in one shot.
任务+背景：我是【某类公司】的【角色】，正在策划【活动/项目】，
受众是【目标人群】。请给【时间线 / N 个创意 / 宣传点子】。
迭代：补充新约束（如"要吸引全球年轻玩家"）→ AI 给出更贴合的版本。
4.2 Tables & Trackers
Why: Many people are intimidated by spreadsheet cells and formulas. AI can generate shift schedules / progress trackers in seconds using natural language, even with complex constraints.
任务+格式+背景：
我有 10 名员工（编号 1-10）。创建一个跟踪每周排班的表格，
列为：日期、员工编号、班次（morning/afternoon）。
约束：每个班次固定 2 人；前一天晚班的次日不排早班；
同一天不排两班；每人每周总班次大致相同。
Iron rule of data privacy: Use employee IDs instead of names, do not enter sensitive information; check company policy before use.
4.3 Meeting Notes
Why: Meetings have lots of information, and task assignments are easy to mix up. Use Google Meet to produce a transcript → feed it to AI for structured organization.
为这次会议生成会议纪要：
- 根据关键要点分章节（而不是简单按发言人流水账）
- 每个要点标注对应发言人
- 把每项 action item 分配给最适合完成的人
The first version cut too finely by speaker; iterating to “cut by key points + label the responsible person” will be much clearer.
4.4 Data Visualization
Why: Presenting data means “letting charts speak” rather than piling up text. AI can help you choose the right chart type and explain why it fits.
背景：我的数据集是一个表格，列有：书名、出版时间、类型、首版时间、销量（百万）。
任务：给几种图表方案，展示"类型"与"销量"的相关关系，并解释每种为何合适。
进阶（模仿参考图）：上传一张范例图 → 先让 AI 描述这张图展示什么 →
再贴自己的数据列 → "建议如何改造这张图适配我的数据，并说明在 Google Sheets 怎么实现"。
4.5 Image Generation
Why: When moving from text generation to image generation, prompt language must change — you must provide specific, vivid visual details.
任务：为海报生成一张电吉他的图像，摄影风格。
细节（缺一不可）：尺寸 / 颜色（闪闪发光）/ 位置（前景，漂浮在天空）/ 整体美感。
迭代：喜欢主体后再加戏 → "把天空变成暴风雨，让闪电击中吉他"。
风格可选：photography / vector art / sketches / watercolors。
For advanced techniques see §5.3 referencing only one dimension of an image; for more product-grade structured JSON text-to-image methods see newtype · Prompt / Context Engineering in Practice §9.
4.6 Speaker Notes
Why: Communication is “telling a story,” not “reading slides aloud.” AI can help you build presentation structure + write speaker notes + source accompanying images.
角色+背景：我是【某品牌】产品设计师。【这里粘贴研究/分析内容】
任务+格式：我该如何组织演示？列出每页幻灯片的主题、关键要点、视觉元素。
（结构通常含：标题页 / 背景介绍页 / N 个功能段 / 行动号召 call to action）
进阶（按听众调）：为这页提供面向【市场总监】的替代讲者备注，
该总监重视精准表达，目标是把握行业趋势 + 建立忠诚用户社区。
练习反馈：我要向经理做演示，想显得权威自信。我会朗读演示稿，
之后请反馈我是否达到目标（节奏 pacing / 语气 tone / 表达清晰度）。
4.7 Other common playbooks (same framework, different fields)
- Writing emails: persona (gym manager) + context (new class schedule available) + task (notify staff) + details (Mon/Wed/Fri class times changed) + tone (professional, friendly, brief). The advantage is keep the information the same, switch audience with one click — rewrite the same email for members, adjusting the tone to “fun + weightlifting puns.”
- Summarizing long texts:
以大纲形式总结这份报告，只保留最核心要点. A makeshift method when the tool doesn’t support long text:summarize by section → then summarize those summaries. You can also change tone by audience (为不擅长数字的人总结/用 80 年代电影台词风格总结). - Tone and style: Treat AI as an actor and the prompt as a script. Use specific words instead of vague ones —
witty banter(witty and teasing) is better thanhumorous,like talking to a friendwhich is better thancasual. You can provide style reference samples, or add constraints (避免行话). - Multimodal prompts: Provide image + text / audio + text at the same time to help the AI understand more accurately and reduce rework. For example, upload a nail art photo along with "write a fun, short social media promotional post"; upload a meeting schedule image along with "extract the speakers and the times of the two panels into a table", then after getting the table follow up with "draft a reminder email based on this".
- Decoding formulas / jargon / error messages:
用分步骤简单解释这个公式 [粘贴公式];用简单话解释什么是复利,假设我完全不懂金融,给个带数字的例子;"Error 404" 是什么意思,列两个常见原因. The AI acts like a personal translator, making jargon, unfamiliar formats, and foreign languages understandable. - Meta-prompting(having the AI help you write prompts, also known as automatic prompt engineering):Generation type——
生成一个能帮我写录用信的提示词/给我一个提示词模板;Optimization type——我该怎么改这个提示，让广告语言更吸引人？(leveling up), remixing multiple prompts into one super prompt, and style swapping to change tone and mood. Meta-prompting acts like a roadmap, guiding you to build more effective prompts step by step.
5. Advanced Techniques and Data Analysis
5.1 Data Analysis Strategies
Why: AI excels at discovering patterns, finding relationships, and identifying trends in large datasets, but prompts must be designed around its capability boundaries, and you must stay alert for hallucinations.
AIs have different strengths with different data types; design prompts according to type: text analysis (topic/sentiment/classification), data augmentation (simulating samples to supplement small datasets,you must clearly label which entries are simulated), Q&A analysis, scenario analysis ("what happens if customer traffic increases by 20%"), image analysis, and customer market research.Complex statistics don't ask it to calculate directly; instead, ask it to provide methods and steps.
Two iron rules:
- Prevent hallucinations: When generating 100 data entries, only a few may be wrong,spot checks are not enough; try to generate information that can be cross-validated.
- Prior data quality: Check 6 items before input — up-to-date / accurate / complete / consistent / relevant / unbiased. You can ask the AI to help check
识别错误、不一致、缺失值，并建议如何清洗标准化. - Encourage exploration: To get the AI to discover hidden trends, use open-ended questions instead of overly detailed instructions (e.g., "You are a journalist investigating real estate trends; how would you tell a compelling story using this vacant housing data?").
5.2 Personal Prompt Library + Prototype AI Agents
- Build a personal prompt library: Useful prompts should be named and saved for later reuse. Don't discard effective prompts; just modify the task/background/references on top of them (like a weekly family menu — don't plan from scratch, tweak slightly based on taste). The same information source can be adapted into multiple versions by switching personas (consumer FAQ → enterprise FAQ).
- Design a prototype AI agent(from Key takeaways): Assign a role + provide scenario context + specify the conversation type + set a stop phrase + have it summarize takeaways and improvement points after ending. Can be used to practice skills, simulate collaboration, and get feedback.
- Communities:
promptingguide.ai(systematic teaching of prompting),huggingface.co(models/datasets/prompt resources).
5.3 Reference Only One Dimension of an Image (to prevent "copying the entire image")
Why: When you give an entire image as reference, models often copy over the subject, composition, and elements entirely. The solution is to narrow "reference this image" down to "only reference a specific dimension of this image", and explicitly state allowed transfers vs prohibited transfers.
任务：生成一张全新插画，主题【新主题】。
参考范围：参考上传图，但只提取【配色 / 笔触质感 / 造型语言】其一。
允许迁移（Do）：如配色——主色2+辅色2+点缀色1、冷暖、饱和度、明度对比。
禁止迁移（Don't）：任何物体、人物姿势、符号、构图布局、可辨识轮廓。
生成前自检：先用文字列出参数（如 5 色调色板），再说明如何应用到新主题。
- A two-step method is more reliable: First have the AI act as a "style analyst" and only analyze the reference image, output a dimension checklist + negative list → you select Do/Don't items → then generate.
- Break free from being "hijacked by the previous image": When the difference between the new image and the old image is too small, add
Ignore all previous generated results. Do not modify based on previous images. Redesign a completely new version.
6. Using AI Responsibly (human-in-the-loop)
Why this is a must-read: AI does not have genuine critical thinking, does not understand nuanced context, and will not bear consequences for you.Responsibility always lies with the person using it. This section is a checklist you can directly use when briefing teams or clients.
Core principle: human in the loop — AI outputs must be verified and judged by a human before being used.
| Risks | Manifestations | How to prevent |
|---|---|---|
| Hallucination | Producing inconsistent / incorrect / meaningless content ("guessing answers" when instructions are too vague) | Fact-check critical content + cross-validate; provide sources / consult experts; express yourself more clearly (false premises will mislead, e.g., "why is Toronto the capital of Canada") |
| Bias / Stereotypes | Linking occupations/activities to specific ages, genders, or cultures | Explicitly require diversity(来自世界各地的美食); point out issues in the next round and request corrections; use inclusive language(service person rather than serviceman) |
| Data security | Inputting confidential / private information into tools | Check company policies before input; prioritize enterprise-grade tools; understand how data will be used |
| Irrelevance / Misunderstanding | Taking idioms literally (e.g., think outside the box) | Provide references; switch to clearer expressions(考虑独特、非常规的方法) |
Workplace checklist (5 items):
- Is this task suitable for using AI? Could it reinforce bias?
- Before using it on company/client projects,have you obtained approval from the relevant responsible person?
- Privacy and security: Is the tool secure? Is there an enterprise version?
- Before use,evaluate output accuracy.
- Disclose AI usage to teams/clients and maintain transparency.
Pro tip: When unsure,clear memory / start a new conversation — this protects privacy, stops bias carryover, helps the AI focus on the current task, and troubleshoots abnormal outputs.
7. Anti-pattern Quick Reference
| ❌ | ✅ |
|---|---|
| Instructions are too vague ("give me five ideas") | Apply TCREI: Role + Audience + Format + Context + Reference |
| Starting over from scratch when the first version doesn't work | Iterate in small steps with ABI, modifying using one of the 4 methods |
| Cramming 3 tasks into one prompt | Split into a prompt chain, one step per task |
| Treating "prompt chaining" as "repeatedly revising the same sentence" | A chain = using the previous step's output as the next step's input |
| Asking the AI to directly give the shortest summary | Compress in tiers via chain of density, checking that no information is lost / nothing is fabricated |
| Writing only one vague description for text-to-image | Provide size/color/position/style; if referencing an image, clearly state Do/Don't |
| Entering employees' real names directly into tables | Use numbered identifiers for desensitization, check company policies first |
| Using AI output directly as-is | Apply human-in-the-loop; evaluate accuracy before use |
| Can't figure out how to write a prompt | Meta-prompting: have the AI help you generate / optimize the prompt |
| Forgetting good prompts after using them | Name them and save them in your personal prompt library, reuse by modifying fields next time |
References and Extensions
Source: Notion notes for all 30 lessons of Google's "Prompting Essentials Specialization" + parent intro (mixed Chinese-English, full_scan). Covers: 5-step framework / hands-on examples / 4 iteration methods / multimodality / text-to-image / referencing only dimensions / responsible use / emails / personalized tone / brainstorming / tables / Gemini Workspace / meeting minutes / summarization (including chain of density) / data analysis and discovery / decoding tables and everything / visualization / presentation notes and feedback / prompt chaining (discovery + multi-step) / meta-prompting / AI agents / Key takeaways.
Companion wiki:
- newtype · Prompt / Context Engineering in Practice — Lao Huang's "prompt engineering doesn't need to be studied" stance + Prompt House / Ultra think / structured JSON text-to-image / Skill end state; the product-grade and counter-consensus complementary side of this article
- newtype Prompt Template Library (4 production-grade prompts + 1 CLAUDE.md sample) — Reusable prompt template collection
History
- 2026-06-03 v1.0 created — Based on full_scan notes for all 30 lessons of Google Prompting Essentials under the will-ai-note project. Organized into 7 sections: 5-element TCREI framework / 4 iteration methods / prompt chaining (including CoT·ToT variants) / 6+N high-frequency office scenario playbooks / advanced topics and data analysis / responsible use checklist / anti-patterns. Cross-referenced with newtype-prompt-engineering, positioning Google's official standard framework (ready for beginners) against Lao Huang's counter-consensus stance (product-grade orchestration).