Module 4 is the hands-on module. Where Module 3 mapped the landscape, this one picks up the tools a working professional actually touches: how to write a prompt that behaves predictably, and which Gemini surface to reach for. Most of the exam weight here is discrimination between products that sound alike — Gems, Saved Info, Gemini Advanced, and Gemini Notebook all “personalize” something, and the questions turn on which one personalizes what.1
A caveat on the exam notes in this module. The deck’s answer slides marked correct options graphically, and that highlighting didn’t survive text extraction. The answers below are derived from the deck’s own definitions elsewhere — high-confidence, but not verbatim-confirmed the way the other modules’ are.
Prompting by example: zero-, one-, and few-shot
The three techniques differ by exactly one variable — how many worked examples you put in the prompt before the real request. More examples means the model has more to pattern-match against, which is why reliability climbs across the three.
- Zero-shot — ask the model to complete a task with no examples. It answers purely from what it already knows. “What is Cymbal Retail’s returns policy?” — the model answers from general knowledge, having no company-specific training, which is exactly the weakness of this approach for company-specific tasks.
- One-shot — show the model one example, then give it the real task. A single customer inquiry about shipping times paired with an ideal response, then a new query — the output follows that example’s pattern.
- Few-shot — provide multiple examples. Several inquiry/response pairs covering shipping and returns, then a new query — the output matches the desired pattern more reliably than one-shot does.
Few-shot is the most reliable of the three when you need output in a specific format or voice — the deck makes that comparison explicitly against one-shot.
Exam note: the technique relying entirely on the model’s pre-existing knowledge, with no examples provided, is zero-shot. “Multi-shot” appears as a distractor and is not one of the course’s terms.
Role prompting
Role prompting assigns the model a specific role or persona to guide its behavior. It works because a persona carries a whole bundle of implied conventions — vocabulary, level of formality, what counts as relevant — that you’d otherwise have to specify one instruction at a time. What it controls: style, tone, and focus.
- “You are a customer service agent for a telecommunications company with ten years of experience. A customer is inquiring about their latest bill, which is higher than expected.” → a helpful, patient response that explains possible reasons, offers solutions, and states policy clearly.
- “You are a marketing copywriter for a new line of athletic wear. Write a product description for a pair of running shoes emphasizing comfort and performance.” → persuasive, audience-targeted copy.
The template the course gives: “You are a [persona]. Please write a [task].”
Exam note: the main advantage of role prompting is that it influences the style, tone, and focus of responses by assigning a persona — not that it helps the model learn from examples (that’s few-shot) or understand complex questions. And in the language-tutor question, the best role prompt is the one supplying both a persona and a manner — “a helpful and encouraging tutor who provides constructive feedback” — over one that just names a function.
Prompt chaining
Prompt chaining is a multi-turn sequence in the same thread where each prompt builds on the model’s previous response, converging on a more sophisticated result than a single prompt would produce. In Gemini you keep the thread going without repeating history, and you can name and pin chats.
Four steps:
- Set a goal — a multi-step task or complex question. A business proposal, a trip itinerary, a multi-element marketing image.
- Start the chain — an initial prompt that sets the stage. “I need to plan a 5-day business trip to Tokyo in August. Can you help me create a preliminary itinerary?”
- Build the chain — follow-up prompts that refine. “Can you suggest hotels near the conference center?” then “Can you add day trips to nearby cities?”
- Observe and analyze — watch how each prompt compounds on the last.
Worked scenario: a marketing manager first asks for anti-aging organic skincare campaign ideas and gets social, influencer, and content directions; the follow-up narrows that into a three-month content calendar for the same audience.
Reusable prompt templates
A template is a prompt written once with [bracketed] parameters, then reused
across a role’s recurring tasks. The value is consistency — the same structure
produces comparable output every time, and colleagues can share a template
rather than each rediscovering what works.
- Marketing manager — “Write a catchy social media post for
[platform]promoting our new[product/service]. Highlight[key features]and include a call to action to[desired action].” - Content writer — “Write a
[content type]about[topic]. The target audience is[audience]. The tone should be[tone]. Include[keywords].” - Customer support — “Respond to this customer email:
[email]. Be empathetic and address their concerns about[issue]. Offer a solution that aligns with our[company policy].” - Software developer — “Write Python code to
[task]. Use[libraries/frameworks]and follow[coding style guidelines].”
Grounding and RAG
Grounding is the model’s ability to connect its output to verifiable, specific sources of information — the direct remedy for hallucination, covered in full alongside the other limitation fixes in Module 2.
RAG (retrieval-augmented generation) is grounding implemented through search. This module frames it as two steps: retrieve relevant information from a vast knowledge base, then generate the final output using what was retrieved. The architectural difference is visible in the flow:
- Without RAG: Prompt → Model → Output
- With RAG: Prompt → Query → Vector DBs → Model → Output
Three benefits:
- Improved accuracy and relevance — outputs are more accurate and informative.
- Improved explainability and transparency — the system can show which specific sources produced the output.
- Extended capabilities — the model responds based on context it was given, reaching beyond what it memorized during training.
Module 5 covers the full three-stage version (retrieval, augmentation, generation) and the retrieval tooling.
Exam note: RAG’s primary function is enabling the model to access and use external knowledge sources when generating outputs. The distractor to avoid is “enhance the model’s ability to memorize” — RAG is the opposite of memorization, which is precisely its point.
Google Workspace with Gemini
Google Workspace is the suite — Gmail, Calendar, Drive, Meet, Chat, Docs, Sheets, Slides, Forms, Keep, Voice, Sites, Tasks, Apps Script, Admin. Adding Gemini puts AI inside the apps people already work in, which is the whole value proposition: no context switch, and the AI can see your existing material.
The Gemini side panel appears alongside many Workspace apps and can summarize, analyze, and generate content using insights from your emails and documents — without switching applications or tabs. A new employee uses the Drive side panel to get up to speed on existing projects.
Per-app capabilities:
| App | What Gemini does | Example |
|---|---|---|
| Gmail | Write and refine emails | An HR rep drafts a company-wide benefits announcement |
| Docs | Content creation, refining, proofreading in-document | A marketer generates taglines and social posts for a shoe campaign |
| Slides | Generates photorealistic images for decks | ”Create an image of a modern office building” |
| Sheets | Generates formulas from a natural-language description | ”Create a formula to find cell C1 in range D:G and output the value in column G” |
| Meet | Real-time transcription and translation; meeting summaries and action items | Translating a meeting into Spanish live |
Two further Workspace tools:
- Google Vids — online video creation and editing; generates a first draft of a video for project updates, timelines, and insights.
- AppSheet — no-code app development; creates an app structure with tables, columns, and links that you then review and edit.
Exam note: for a multilingual meeting where attendees face language barriers, the answer is live translated captions with speaker identification — real-time, in-meeting. Post-hoc translated summaries in the chat are the near-miss distractor.
The Gemini surfaces: Advanced, Saved Info, Gems, Notebook
This is the module’s highest-yield section. Four products, all of which “customize” Gemini in some sense, and the exam gives you a scenario and asks which. The organizing question: what is being personalized — the model tier, a persistent fact about you, a reusable assistant, or a closed set of sources?
- Gemini Advanced — the upgrade tier of the base Gemini app. Strongest at coding, logical reasoning, following nuanced instructions, and creative collaboration, with adjustable retention settings and priority access to the newest experimental features and models. You pick it when the task is hard, not when you need customization.
- Saved Info — a setting that lets Gemini remember specific context without carrying the entire conversation history. A sales rep tells Gemini once that they’re a sales rep, so that context persists — but they don’t want feedback on a new pitch colored by earlier practice sessions.
- Gems — personalized AI assistants inside Gemini, pre-made or custom, for repeatable tasks and consistent interaction styles. Three properties: personalized responses (tailored instructions increase relevance), streamlined workflows (saved templates and prompts eliminate repetition), and reset context — each chat with a Gem is isolated, like separate conversations with the same expert. You build one by identifying a base Gem (considering functionality, license, and simplicity), copying it, and then supplying role-prompting, documents, clear instructions, and examples.
- Gemini Notebook (NotebookLM) — a research partner grounded only in sources you upload, which is what makes every answer traceable. It does not draw on general web knowledge.
Don’t confuse these: Saved Info vs. Gems
Both persist something across sessions, and they’re the most confusable pair in the module. Saved Info retains light context about you and carries it forward. A Gem resets context on every chat. So a scenario asking for feedback uninfluenced by previous sessions wants a Gem, not Saved Info — the isolation is the feature being tested.
Gemini Notebook in detail
- Source types: Google Docs, Google Slides, YouTube videos, websites, audio files, PDFs, text files. Notebook extracts text and saves a local copy; it cannot delete or edit your original files.
- Access: Viewer or Editor roles. Sharing a notebook does not share the underlying source file, and chat-only mode hides the sources entirely.
- Sources and sync: size and usage limits apply, out-of-date sources must be re-synced manually, and notebook data is not used to train models.
- Configure Chat: conversation style (Default, Analyst, Guide, Custom) and response length (Default, Longer, Shorter).
- Studio → Audio Overview: a two-host “Deep Dive conversation” about your sources (English only), with an Interactive mode letting you join and ask questions.
- Three advantages over Gemini and Gems: hyper-focused knowledge (only your sources), interactive learning (questions, summaries, quizzes), and source-based answers (every insight traceable to an upload).
- Tiers: Pro adds enhanced research, increased capacity, response length/style customization, and usage analytics. Enterprise builds on Pro with compliance and administrative features, extra privacy and security, sharing restricted to chosen collaborators, and IAM-managed access.
Organizational notebook patterns from the deck: sales (product info, competitor analysis, market research), marketing (blog posts, reports, webinars), training and onboarding (manuals, presentations, videos), development (specs, design docs, feedback), teaching (strategy plans, standards, lecture notes — with Studio outputs like study guides, briefing docs, FAQs, timelines), and customer support (help center articles, product docs, FAQs as a shared knowledge base).
Quick-match cheat sheet
Straight from the deck’s matching activity:
| Need | Tool |
|---|---|
| Create your own custom version of Gemini | Gems |
| Personalize an assistant for a repeatable task or interaction style | Gems |
| Assistant grounded only in documents you provide, no web knowledge | Gemini Notebook |
| Write, summarize, and brainstorm inside Docs, Gmail, and Sheets | Google Workspace with Gemini |
| Priority access to the newest features; best at coding and reasoning | Gemini Advanced |
| Feedback isolated from prior sessions | Gems (reset context), not Saved Info |
Exam note: two more scenarios worth pre-loading. A freelance writer switching between content types who wants consistency → Gems. A product manager prepping for a launch meeting from Drive documents → Gemini Notebook summarizes key findings, identifies challenges, and provides talking points — it does not auto-build a slide deck or send calendar invites, both of which appear as distractors. And on grounding: it connects responses to specific sources and is not exclusive to Notebook; Notebook can generate quizzes from your documents but cannot access the whole internet.
Gemini for Google Cloud
The developer- and admin-facing line. Each product is Gemini embedded in one Google Cloud surface, and the exam asks which surface fits a stated job.
- Gemini Cloud Assist — design, manage, and optimize applications on Google Cloud; personalized guidance and lifecycle management; analyzes your cloud environment, deployed resources, metrics, and logs.
- Gemini in BigQuery — makes data analysis more accessible: writes code, helps you understand your data, generates insights automatically.
- Gemini Code Assist — an AI pair programmer offering code suggestions, generated blocks, and explanations across 20+ languages, editors, and developer platforms.
- Gemini in Colab Enterprise — AI assistance in notebooks, suggesting and generating Python from your descriptions.
- Gemini in Databases — helps developers and DBAs manage databases, simplifying many aspects of database work.
- Gemini in Looker — analyze data, create visualizations, generate reports.
- Gemini in Security (Security Command Center) — helps security teams detect, contain, and stop threats; near-instant analysis of security findings and potential attack paths; summarizes threat-actor tactics, techniques, and procedures.
Enterprise security: prompts and responses to Gemini for Google Cloud are not used for training, and standard Google Cloud protections apply.
Exam note: a development team wanting better coding efficiency and collaboration → Gemini Code Assist. Cloud Assist is the distractor; it manages infrastructure and applications, not the act of writing code.