If you've ever typed a prompt into ChatGPT or Copilot and gotten back something generic, the problem usually isn't the AI — it's the framework behind your prompt. I recently went through an AI Brainstorming and Prompt Engineering course and answered a long list of scenario-based questions along the way. Instead of letting those notes sit in a document, I've turned them into this guide.
Below you'll find a concise theory section covering the AI Brainstorming Framework (Diverge, Deepen, Decide), the CROSS prompting framework, how to catch AI hallucinations before they cause damage, chained prompting for complex tasks, and how businesses should think about Generative AI APIs. After the theory, I've included the full FAQ — all 30+ practice questions and answers from the course, organized by topic.
Table of Contents
- The AI Brainstorming Framework: Diverge, Deepen, Decide
- The CROSS Framework for Prompt Engineering
- Trust But Verify: Hallucinations, the Verification Pyramid, and the "Confident Impostor"
- Chained Prompting: Breaking Big Problems Into Small Steps
- Generative AI APIs: The Business View
- FAQ: 30+ Practice Questions & Answers
1. The AI Brainstorming Framework: Diverge, Deepen, Decide
Most AI-assisted brainstorming sessions fail for the same reason unstructured human brainstorming fails — there's no process. A simple three-phase framework fixes this:
- Diverge: The goal here is volume and variety, not quality. You ask the AI to generate as many different ideas as possible, even unconventional ones. Judging ideas too early kills creativity.
- Deepen: Now you take the most promising ideas and flesh them out — core functionality, target user, key benefit, and success metric. This is where a well-structured prompt (persona + context + objective) turns a one-line idea into a usable proposal.
- Decide: Finally, you need to choose. For a resource-constrained project, an Impact vs. Effort Matrix is usually the fastest way to prioritize, since it plots ideas against how much value they create versus how hard they are to build. For final shortlists, a decision matrix lets you score each option objectively against your actual business criteria instead of going with gut feeling.
The same divide-and-conquer thinking applies to multiple personas: instead of asking one AI persona for names or ideas, ask it to think as a marketer, an engineer, and a strategist in turn. Each lens surfaces ideas the others would miss.
2. The CROSS Framework for Prompt Engineering
A vague prompt gets a vague answer. The CROSS framework is a checklist for writing prompts that consistently produce useful output. It stands for:
- C – Context: The background the AI needs — what's happening, who the competitor is, what data it's working with.
- R – Role: The persona you want the AI to adopt (e.g., "a seasoned market research analyst with 15+ years in HR technology").
- O – Objective: The specific outcome you want — the exact questions the AI needs to answer.
- S – Scope: Boundaries and structure — how long, how many sections, what to include or exclude.
- S – Source: What the AI should base its answer on (a prior output, a specific document, or "current market trends" if it's a plausible analysis).
A prompt like "Create a post for our Twitter account announcing our new product" is missing Context, Scope, and Source — which is exactly why it produces something generic. Once you've got a first draft, don't restart from scratch. Use iterative refinement: specific follow-ups like "change the opening line to focus on the customer's pain point" move the draft forward far more efficiently than regenerating it.
Output format matters too. Explicitly asking for a markdown table, JSON object, or bullet list isn't about creativity — it's about making the output predictable and machine-readable so it plugs directly into whatever comes next in your workflow.
3. Trust But Verify: Hallucinations, the Verification Pyramid, and the "Confident Impostor"
The single biggest risk in using AI for research or analysis isn't that it's slow — it's that it can be confidently wrong. This is called the Confident Impostor phenomenon: AI presents fabricated or inaccurate information with the same fluent, self-assured tone as accurate information, so there's no obvious signal telling you to double-check.
It helps to separate two distinct failure modes:
- Hallucination: a complete fabrication — a source, statistic, or fact that doesn't exist at all.
- Factual error: an incorrect detail about something real — a wrong date or number attached to a real event.
The Verification Pyramid gives you a way to scale your scrutiny to the stakes involved. At the base you have quick internal gut checks; further up, secondary sources and peer review; and at the top, Primary Source Confirmation — the most rigorous level, reserved for critical, high-stakes claims you can't afford to get wrong.
This is also why a "sourcing mandate" belongs in any AI research prompt: instructing the AI to cite where information comes from doesn't make the AI infallible, but it fulfills your responsibility as the user to actually go verify the claims before you act on them. The underlying principle is Trust but Verify — if an AI analysis surfaces a surprising insight, the right next step is to manually check a sample of the underlying source material, not to forward it straight to leadership.
4. Chained Prompting: Breaking Big Problems Into Small Steps
Complex deliverables — a full report, a multi-part analysis — rarely come out well from a single mega-prompt. Chained prompting solves this by splitting the work into a sequence of smaller, connected prompts, where each step's output feeds the next.
The first phase, Deconstruct, is where you break a large, complex problem into a series of smaller, logical, manageable subtasks before you ask the AI to do anything. This mirrors the Four-Step EDA Methodology (Orient → Visualize → Correlate → Hypothesize) used for exploratory data analysis on qualitative feedback: you orient with a high-level thematic summary first, correlate to find patterns across different feedback sources, and only then move to drafting. I've covered that methodology in full detail in a dedicated EDA walkthrough if you want the deep dive.
When it's time to turn analysis into a document, the single most important thing to give the AI isn't a word count or a tone instruction — it's a clear, logical outline to follow. Everything else (tone, length) is easy to refine afterward; structure is not.
5. Generative AI APIs: The Business View
Once you move from chatting with an AI tool to integrating a Generative AI model into a product, you're working through an API. The mental model that makes this click: the AI model or service is the powerful, complex kitchen; the API is the standardized way any application places an order with that kitchen without needing to know how to cook.
From a business standpoint, three things matter most:
- Cost: using a provider's API gives you access to advanced models without the capital cost of building one internally.
- Access control: every application should get its own unique API key. This isolates access and cost per application and allows targeted revocation if one key is ever compromised — reusing one company-wide key across projects defeats this entirely.
- Stability and budget: rate limits exist to manage server load and keep the service stable for everyone, while hard spending limits are how you control operational budget and financial risk, independent of performance.
Structurally, every API call has a request (your input data plus instructions) and the AI model's response — the same request/response pattern behind every Generative AI integration, from a chatbot to a document summarizer.
FAQ: 30+ Practice Questions & Answers
Here's the full set of scenario-based questions from the course, with every answer option, the correct choice, and a short explanation of why — grouped by topic.
Brainstorming & Idea Prioritization
Q1: Your goal is to understand the key strategic takeaways from three lengthy market research reports. Which of the following prompts is the most effective?
- Instruct the AI to summarize each of the three reports individually.
- Ask the AI to synthesize the most important strategic themes that are common across all three reports. ✅
- Prompt the AI to check if the information in the three reports is interesting or relevant.
- Command the AI to combine the text from all three reports into a single document.
Why: Individual summaries or a combined document still leave you to do the analytical work yourself. Only synthesis across all three sources surfaces the shared strategic themes, which is the actual insight you're after.
Q2: A team is tasked with generating ideas for a near-term project with a limited budget. Which option is most appropriate for prioritizing their ideas?
- Impact vs. Effort Matrix ✅
- Value vs. Novelty Matrix
- SWOT Analysis
- Beginner's Guide structure
Why: An Impact vs. Effort Matrix plots ideas directly against business value and implementation cost — exactly the two variables that matter under budget and time constraints. SWOT is a positioning tool, not a prioritization tool.
Q3: What is the primary goal of the "Diverge" phase in the AI Brainstorming Framework?
- To select the single best idea for immediate implementation
- To analyze the pros and cons of a few promising options
- To generate a high volume and wide variety of potential ideas ✅
- To create a detailed, step-by-step plan for project execution
Why: Diverge is purely a generation phase. Evaluating, narrowing, and planning all happen in later phases (Deepen and Decide) — judging ideas too early during Diverge limits creative range.
Q4: What is the primary benefit of using a "format-agnostic" prompting strategy?
- Generating a more creative and unexpected range of ideas.
- Allowing for easy repurposing of core content for multiple formats. ✅
- Ensuring the final output will be 100% factually accurate.
- Enabling the generation of responses longer than a single paragraph.
Why: When core content isn't locked to one output shape, you can reshape the same underlying ideas into a blog post, a social caption, or a slide deck without regenerating the substance each time.
Q5: You need the AI to brainstorm a list of potential names for a new software product. To get the most creative and diverse set of ideas, which strategy is most effective?
- Providing the AI with a list of 50 keywords and a specific tone (e.g., "futuristic, trustworthy") and asking for a combined list.
- Instructing the AI to adopt multiple, distinct personas (like a marketer, an engineer, and a strategist) and generate ideas from each perspective. ✅
- Using self-consistency prompting to generate three separate lists and then asking the AI to select the best 10 from the combined pool.
- Restricting the AI to names that are only two syllables long and do not contain vowels.
Why: Each persona brings a genuinely different lens (marketing appeal vs. technical accuracy vs. strategic positioning), which produces more varied ideas than a single perspective working from a keyword list, and more than the AI judging its own output.
Q6: A manager provides an AI with a list of brainstormed ideas and asks it to "develop a proposal outline." This is an example of using AI for which primary task?
- Generating a set of new and creative marketing ideas
- Summarizing a single, lengthy source document
- Transforming unstructured ideas into a logical document structure ✅
- Verifying the factual accuracy of the brainstormed ideas
Why: Turning a raw idea list into an outline isn't generation, summarization, or verification — it's imposing logical structure on unstructured input.
Q7: In the "Diverge, Deepen, Decide" framework, what is the main purpose of using a decision matrix in the final phase?
- To generate more creative ideas when the brainstorming session has stalled
- To explore the detailed pros and cons of the single most popular idea
- To create an objective comparison of the top options against key business criteria ✅
- To check the AI's initial research for any potential factual errors
Why: By the Decide phase you already have a shortlist; a decision matrix scores each option against the same weighted criteria so the choice is objective rather than based on gut feeling.
Q8: What is the primary strategic benefit of using an AI to generate "what if" scenarios for a business plan?
- To get a definitive and accurate prediction of future company revenue
- To create a complete list of all possible things that could ever go wrong
- To summarize the key points and takeaways of the existing business plan
- To challenge assumptions and build a more resilient, future-proof strategy ✅
Why: "What if" scenarios aren't predictions — the AI can't guarantee future outcomes. Their value is stress-testing the assumptions your plan currently relies on, so the strategy holds up under different conditions.
The CROSS Prompting Framework
Q9: Which of the following are components of the CROSS framework for prompt engineering? (Choose five)
- Role ✅
- Context ✅
- Bias
- Scope ✅
- Source ✅
- Creativity
- Hallucination
- Objective ✅
- Nuance
- Verification
Why: CROSS is an acronym for Context, Role, Objective, Scope, and Source. Bias, creativity, hallucination, nuance, and verification are all real AI concepts, but none of them are part of this specific prompting checklist.
Q10: Your team needs to generate a social media post. You use the following prompt: "Act as our social media manager. Create a post for our company's Twitter account to announce our new product." Which elements of the CROSS framework are clearly missing from this prompt? (Choose three)
- Role
- Context ✅
- Scope ✅
- Source ✅
- Objective
- Tone
- Audience
- Format
Why: The prompt already supplies a Role ("social media manager") and an Objective (announce the product). What's missing is background information about the product (Context), boundaries like length or hashtags (Scope), and what the post should be based on (Source).
Q11: What is the primary reason for specifying the output format (for example, "in a markdown table," "as a JSON object") in a prompt?
- To make the AI focus on data extraction over creative interpretation.
- To ensure the output is predictable and machine-readable for use in other destinations. ✅
- To train the AI on new, proprietary data structures for future tasks.
- To increase the AI's processing speed and reduce token usage for cost savings.
Why: A specified format guarantees the output can be parsed or dropped directly into another system (a spreadsheet, a database, another script) without manual reformatting.
Q12: You are using iterative refinement to improve an AI-generated draft of a customer-facing sales pitch. Which of the following are good examples of effective follow-up prompts for guiding the AI toward a better final version? (Choose three)
- "Ensure the final call-to-action is in bold and is an active hyperlink."
- "Make the pitch sound more compelling and dynamic." ✅
- "Change the opening line to focus on the customer's pain point rather than our product features." ✅
- "Rerun the prompt to get a completely new version from scratch."
- "Make the tone more optimistic." ✅
- "Rewrite the second paragraph."
Why: Effective iterative refinement gives the AI clear, specific creative direction that builds on the existing draft. Restarting from scratch throws away a usable draft, while "rewrite the second paragraph" gives no direction on what should actually change.
Q13: A manager needs to generate a summary of a technical project update for a non-technical executive. Which prompt is the best example of providing a clear persona and context?
- "Summarize the technical findings in five bullet points and explain the next quarter's steps."
- "Provide a summary of the project's milestones, and use clear, corporate language."
- "Break down the project's strategic value and business impact, acting as the VP of Strategy." ✅
- "Act as an independent financial analyst and extract the three largest risks and the final ROI projection."
Why: This is the only option that assigns an explicit persona (VP of Strategy) tied to a specific lens (strategic value and business impact) — exactly what a non-technical executive audience needs. The financial analyst option has a persona too, but the wrong one for this audience and goal.
Trust, Verification & Hallucinations
Q14: According to the Verification Pyramid, which level of verification is the most rigorous and should be used for the most critical, high-stakes information?
- Level 1: Internal Gut Check
- Level 2: Secondary Source Verification
- Level 3: Primary Source Confirmation ✅
- Level 4: External Peer Review
Why: Going directly to the original, primary source is the strongest available check — stronger than a quick internal sanity check or relying on a secondary source that has already filtered or reinterpreted the information.
Q15: What is the primary difference between a hallucination and a factual error in AI-generated content?
- A hallucination is a visual error, while a factual error is a text-based error.
- A hallucination is an incorrect detail about a real thing, while a factual error is a complete fabrication.
- A hallucination is a complete fabrication, while a factual error is an incorrect detail about a real thing. ✅
- A hallucination is a biased statement, while a factual error is an outdated statistic.
Why: The distinction matters because they need different fixes: a hallucination (an invented source or event) has to be removed entirely, while a factual error (a wrong date attached to a real event) just needs correcting.
Q16: Which of the following best defines the "Confident Impostor" phenomenon in AI?
- An AI model that uses confident language to hide its inability to complete a complex task.
- An AI that presents inaccurate or fabricated information with a high degree of confidence and fluency, making it appear correct. ✅
- An AI that correctly identifies factual errors in its own outputs before they are shared.
- An AI that intentionally adds biased information to its responses to manipulate a user's perception.
Why: The core danger is that fluent, confident-sounding text carries no signal about accuracy — wrong answers read exactly as convincingly as correct ones, so tone can never substitute for verification.
Q17: An AI analysis of customer feedback provides a surprising insight. According to the "Trust but Verify" principle, what is your most critical next step?
- Immediately share the surprising insight with your leadership team
- Ask the AI to rewrite the insight to make it sound more plausible
- Manually review a sample of the source feedback to confirm the AI's finding is accurate ✅
- Discard the insight because it contradicts your own professional expertise
Why: "Trust but Verify" means surprising, high-impact claims specifically warrant manual spot-checking against the underlying source data before you act on or share them — not automatic forwarding, and not automatic dismissal either.
Q18: What is the main purpose of adding a "sourcing mandate" to an AI research prompt?
- To fulfill the user's responsibility to verify claims against real sources ✅
- To ensure the AI uses only the most creative and interesting sources
- To make the AI's final response longer and more detailed
- To help the AI learn where it gets its information from for future use
Why: Asking the AI to cite sources doesn't make it infallible — it gives you something concrete to actually check, which is your responsibility as the person using the output.
Chained Prompting & the EDA Methodology
Q19: When designing a multi-step "chained prompt" sequence, what is the primary purpose of the Deconstruct phase?
- To write the final executive summary by combining all previous outputs.
- To break a large, complex problem down into a series of smaller, logical, and manageable subtasks. ✅
- To assign the AI a specific persona that it will use for the entire workflow.
- To fact-check the output of each step in the chain before proceeding to the next one.
Why: Deconstruct happens before any prompting starts — it's the planning step where you map out the subtasks the chain will work through, rather than a step that assigns personas or checks facts.
Q20: To begin your analysis of a large volume of unstructured customer feedback, which prompt is most effective for the "Orient" step, the first part of the Four-Step EDA Methodology?
- A prompt asking the AI to find the single most important customer quote
- A prompt requesting a high-level thematic summary of the main issues ✅
- A prompt instructing the AI to correlate the feedback with sales data
- A prompt asking the AI to generate three business hypotheses
Why: Orient is about getting your bearings first — a broad thematic overview — before narrowing to a single quote, correlating data sources, or generating hypotheses, which all belong to later steps.
Q21: What is the primary purpose of the "Correlate" step in the EDA methodology for qualitative feedback?
- To get a basic, high-level summary of a single source of feedback
- To turn an AI-generated insight into a testable business hypothesis
- To identify hidden patterns by connecting themes across different feedback sources ✅
- To create a text-based breakdown of the most common complaint themes
Why: Correlate specifically looks across multiple sources to find connections a single-source summary would miss — hypothesis generation comes after this step, not during it.
Q22: When prompting an AI to draft a formal report from your completed analysis, what is the most important component to provide in your prompt?
- The desired word count for the final document
- A clear, logical outline for the AI to follow ✅
- A request for the AI to use a professional and authoritative tone
- A list of all the raw data you used in your analysis
Why: Structure is far harder to fix after the fact than tone or length. An outline anchors the entire draft; tone and word count are easy adjustments once the structure is right.
Q23: What is the primary goal of the "Iterative Refinement" process when working with an AI-generated draft?
- To verify the factual accuracy of all data points in the text
- To ensure the document's structure follows the approved outline
- To expand the draft by generating additional content for each section
- To polish the existing text for tone, clarity, and strategic impact ✅
Why: Iterative refinement is a polishing step, not a fact-checking, structuring, or expansion step — those happen earlier or separately in the workflow.
Q24: What is the primary advantage of using an LLM for initial research on an unfamiliar topic compared to a traditional search engine?
- Gaining access to paywalled or otherwise private information
- Receiving a much longer and more detailed list of websites to visit
- Getting a direct, synthesized explanation of the topic itself ✅
- Ensuring the information provided is always 100% factually accurate
Why: A search engine hands you a list of links to read yourself; an LLM synthesizes an explanation directly — faster orientation, though it still needs the verification steps covered above since it's not guaranteed to be accurate.
Generative AI APIs for Business
Q25: In the video's analogy, what does the restaurant's kitchen represent?
- Your company's existing application (for example, your CRM).
- The complex, powerful AI model or service. ✅
- The end user making a request.
- The API itself.
Why: The kitchen does the heavy lifting out of sight, just like the AI model does the actual computation. The API is the equivalent of the menu/waiter — the standardized way to place an order without needing to know how to cook.
Q26: What is the primary business benefit of using an AI provider's API instead of building an AI model from scratch?
- Guarantees faster response times than a custom-built, internally hosted model.
- Grants full ownership and customization rights over the AI model's architecture.
- Provides access to advanced AI models without the high cost of internal development. ✅
- Eliminates all internal data security risks by outsourcing the AI processing.
Why: Building a competitive foundation model requires enormous R&D investment. An API gives you that capability at a fraction of the cost — though it doesn't guarantee speed, ownership, or eliminate your own security responsibilities.
Q27: A developer on your team suggests using the same powerful, company-wide API key for a new, small-scale experimental project. Which business principle does this violate?
- Limiting API access to specific times of the day.
- Encrypting all data transmitted to the AI model.
- Rotating all API keys at the end of every week.
- Isolating applications with unique keys to control access and cost. ✅
Why: A shared key means you can't track spend or revoke access for one project without breaking every other project using that same key. Unique keys per application keep cost and access cleanly isolated.
Q28: What is the primary purpose of an API "rate limit"?
- To calculate the total monthly cost for billing purposes
- To manage server load and ensure service stability for all users ✅
- To authenticate the identity of the application making the request
- To restrict access to the API based on the user's geographic location
Why: Rate limits are a stability mechanism — they prevent any single application from overwhelming shared infrastructure, protecting service quality for every user on the platform.
Q29: Setting hard spending limits on API usage is a primary method for controlling which business factor?
- The performance and speed of the AI model
- The level of access for different user groups
- The quality and accuracy of AI outputs
- The operational budget and financial risk ✅
Why: Spending caps are a financial control, separate from rate limits (stability) and API keys (access control) — each mechanism manages a different business risk.
Q30: In a typical API call to a Generative AI service, what does the "request" body contain?
- The API key and the user's password
- The complete source code of the AI model
- The input data and the instructions for the task ✅
- The billing information for the transaction
Why: The request body is where your prompt and any input data live — authentication (the API key) is typically handled separately in the request headers, not the body.
Q31: What is the primary function of an API in the context of Generative AI?
- To store and manage the data used for training AI models
- To provide a standardized way for applications to access an AI model ✅
- To install and run the AI model directly on a user's local computer
- To monitor the ethical compliance of all AI-generated content
Why: This is the core definition of an API — a consistent interface any application can use to call the model remotely, without hosting or running the model itself.
Q32: What is the primary security benefit of assigning a unique API key to each application?
- Allowing isolated access control and targeted revocation ✅
- Enabling customized AI responses based on the application
- Ensuring all applications receive the same performance tier
- Securing a volume discount from the AI service provider
Why: If one key is ever compromised, you can revoke that single key without disrupting every other application — the same principle behind Q27 above, viewed from the security angle.
Related Articles
If you found this useful, these related guides go deeper on specific pieces covered above:
- Generative AI Fundamentals Explained for Beginners (With IT & Network Engineering Examples)
- Generative AI MCQ Quiz for Beginners (With Answers & Explanations)
- The AI Four-Step EDA Methodology Explained for Beginners
Prompt engineering frameworks like CROSS, and process frameworks like Diverge-Deepen-Decide, aren't really about the AI — they're about giving yourself a repeatable process so the quality of your output stops depending on luck. Start applying one framework at a time, and iterative refinement will do the rest.
