Four questions per class, answers hidden until you commit

What are the three pillars the entire semester is built on?

Prompt Design, Thinking Design, and Human-AI Relationship, described as a loop where the first half of the semester lives in Prompt Design, Thinking Design is Dom's own method taught later, and Human-AI Relationship is where the capstone group project lives.

What is Tony's rule about how many AI tools to actually master?

Go deep on two or three tools that fit the specific repeated steps of your own workflow, rather than shallow on a hundred tools, because a Swiss-army kit fits none of them precisely.

According to Dom, what should happen when scope creep pushes a designer toward tasks like copywriting?

A designer should stay in their own trained craft. AI can draft passable copy for anyone, but a trained copywriter pulls better copy from it than a designer can, and a trained designer pulls better design from it than anyone else.

Why did Dom cite Delta's 2018 facial-recognition study during the first class?

To show that AI accuracy for Black women was the lowest of any group tested, even after retraining, as evidence that engineers alone cannot be the only ones building AI and that designers must lead because they carry culture and lived experience into the room.

What is the mechanism the peanut butter exercise in class demonstrated?

That adding more context sharpens prediction: when the class typed the word that follows 'peanut butter,' answers varied, but adding 'and ___' made almost everyone converge on 'jelly,' showing the same predictive math whether the predictor is human or machine.

What are the three layers of AI Dom presented, and where does today's AI sit?

ANI (Artificial Narrow Intelligence), AGI (Artificial General Intelligence), and ASI (Artificial Super Intelligence). The class currently plays in ANI, with AGI still years away and ASI, per Dom, 'a theory, not even a hypothesis.'

How did Dom explain what a hallucination actually is?

He compared it to reading a book from the middle instead of the beginning: you would still tell a story, but you'd make things up to fill the gaps you never read. More tokens of real context reduces that guesswork.

Why did the 'spacesuit prompt' about an astronaut refilling soap in space work better than the sharper, more direct prompts?

Because a generic prompt and a more refined but still-generic prompt both returned ordinary ideas, while reframing the same brief through a completely unrelated world (an astronaut on the ISS) produced something nobody else in the category would have reached for. A lateral reframe beat a 'better' version of the same ask.

What do the four letters of Geoff Gibbins's STOP framework stand for?

Stop, Think, Organize, Proceed. It is meant as a pause for moments that deserve scrutiny, not a checklist to run on every single output.

What are Corrix's three measurement buckets for human-AI collaboration?

Results (do you beat the AI working alone), Relationship (the quality of the dialogue, whether you add context and push back), and Resilience (whether you still understand what the model is doing or are losing the underlying skill).

What is Geoff's single most practical habit for evaluating AI output, and how effective did he say it is?

Ask the AI to fact-check its own answer before you read it. Models are far better at spotting their own errors than avoiding them in the first place, and he said it works roughly nine times out of ten.

Why does training an image model on your own brand guidelines still not solve the sea-of-sameness problem, per Dom's caution in class 3?

Because everyone using the same shared base model drifts toward the same look, so training on your own brand guidelines only gets you dilution, not real distinction. This is why Coca-Cola deep-tunes its own foundation model instead of using shared tools like Krea.

What do the four letters of the GOLD framework stand for?

Goal (what you want the AI to do), Output (format, persona, style), Limitations (what to avoid or emphasize), and Data (the context you provide, like PDFs or past chats).

What are the three layers of Sydney's prompting hierarchy, from broadest to narrowest?

Custom instructions (baseline behavior across the whole tool), Scoped instructions (a Project, Gem, or focus chat), and Task prompts (the exact request in the moment). The model follows the most specific and recent instruction under the umbrella of the layers above it.

How is a debate partner set up so it actually pushes back instead of just agreeing?

You explicitly tell it to act as your opponent, push back for a set number of exchanges (three to four), avoid straw-manning your argument, and stay nuanced, because a model defaults to agreement unless you set the terms of the relationship.

Why does Sydney recommend ending a long or complex prompt with 'do you understand what I'm asking of you?'

It forces the model to restate the task back to you, so you can catch a misparse or a mis-pasted detail before the model actually runs the task.

What are the five slots in the image prompt structure taught in class 5?

Format (what kind of visual, like photo or tintype), Subject (the hero), Details (what's happening), Style (the vibe, colors, textures), and Parameters (aspect ratio, character ref, and other technical settings).

What is the difference between a composition reference and a style reference in weighting?

They are weighted as separate dials in a tug-of-war: whichever weight is set highest is what the model obeys. In the tulip demo, Cref 50/Sref 100 caused the AI to copy grain and lighting while composition played a smaller role, and flipping to Cref 100/Sref 50 locked in composition while deprioritizing style.

What are the five steps in the hallucination troubleshooting method before regenerating an image?

Analyze how consistently the hallucination appears, check your references, evaluate your prompt for vague or sticky words, consider the training data, and use a multimodal tool to troubleshoot the meta question.

Why did the phrase 'no horizon line' keep producing a horizon line in the pear still-life example?

Because AI models don't process negatives well and fixate on the words given to them, so describing what you don't want can still pull that exact idea into the frame. The fix was to describe what you did want instead, like a seamless background.

What is the difference between a keyframe and a reference image?

A reference guides palette and mood and may never appear in the final output, while a keyframe is pinned to a timestamp and locks composition, pose, and camera before the AI animates between two of them.

What are the six slots in the video prompt structure?

Format, Subject, Details, Motion (what moves and what the camera does), Style, and Parameters. Motion is the new slot compared to the image prompt structure, and you should say 'static camera' explicitly if you want no movement.

How was character consistency for the woodcut mouse Baxter built before any video was generated?

Midjourney was used to explore the design, Gemini analyzed the illustration style, ChatGPT built a turnaround sheet showing front, side, and back views as the master reference, and five reference angles of white animals were shot up front so the model had no room to invent features. Claude then wrote the render-model-specific prompt for Seedance.

Why does a flat 2D illustration tend to drift toward photographic realism once it starts moving in a video model?

Video models are trained mostly on photographic and 3D motion, so a flat style drifts toward realism as soon as it animates. Sydney called this the 'six-finger era' of video, and the fix is holding the camera static and pre-flighting the prompt in a smart model before generating.

What are the five rows of the AI-Enhanced Workflow Canvas?

Desired outcome ('if I do it well, it will be...'), Barriers ('but first I must...'), AI use case ('it would be great if I had help...'), AI tools ('for this, I could use...'), and My role (how the job shifts, 'now I've become..., with more time for...').

What is Irina's reframe of how AI should change an organization's approach to human workers?

Instead of organizations thinking 'replace one human with AI,' Irina wants everybody keeping their own AI partner and becoming twice as capable, doubling power rather than replacing a person.

What are the three mindsets used to tag any AI use case in the workflow canvas?

Automate (a repetitive task that just runs, like meeting notes), Enhance (AI works with you, like brainstorming 15 campaign territories), and Advise (AI surfaces insight you couldn't reach alone, like predicting a client's objections from their history). One step can carry more than one tag.

Why does Irina keep one notebook or gem per client instead of starting a fresh chat each session?

So she isn't starting from a blank state: the workspace already knows her project, and she doesn't have to re-explain the brief every time. This is the same discipline behind tools like NotebookLM, which answers only from uploaded sources.

What are the six workflow archetypes Seb Saga taught for building node-based workflows in Krea?

Lab Bench (compare models or settings), Factory (one input, same process at scale), Multiplier (one asset, many scenarios), Relay (output feeds the next model), Compositor (assemble the pieces, then unify), and Control Rig (ControlNets and structural prompting).

What six questions should you ask yourself before building a node workflow at all?

Does it need to scale, can it be reused or handed off, are you comparing models, is it repetitive, do you need several models orchestrated together, or do you need absolute control. If the answer to all six is no, a single prompt still does the job.

How did Seb Saga's cost-discipline method work when building a node workflow?

He built and verified the full chain on the cheapest capable model first, such as running a skateboard product photo through cheap models for four style variations, then upgraded only the one node that actually needed a stronger, frontier model.

Why did Seb Saga save his LLM-node system prompt as a reusable preset instead of rewriting it each time?

The preset instructs the model to expand a short image prompt into settings, tone, feel, angle, and color palette, so it can be written once and applied to any image-generating node across every future workflow instead of retyping it each time.

What do the four letters of Tony Jones's GOLD prompt structure stand for?

Goal (what you're actually after), Output (ideas, not a finished piece), Limitations (stay in conversation, nothing locked in on the first pass), and Data (the topic, artists, and styles you feed in).

What are the seven principles of "Wrestle With AI for Concepting" that Tony Jones taught?

Be methodical (stage theme to imagery to verses to chorus to polish), start with the hook, interrogate every idea, generate 3 to 5 variations, change the angle of attack when stuck, draw on named artists, and refine with constraints like rhyme, syllables, and mood words.

How do brackets and parentheses direct the lyrics Suno reads?

Structure tags like [Verse], [Chorus], or [Instrumental] go in brackets, and delivery cues like (whispered) or (building) go in parentheses; Suno reads both as instructions rather than as lyrics to sing.

Why did Tony tell the class to study Suno's own top tracks before writing a single lyric?

To learn the tool's vocabulary first, the same tool-review-before-generating habit the room had already built with images, rather than writing lyrics blind.

What are Margaret Boden's three types of creativity as presented in class 10?

Exploratory (deep immersion in a known style, AI's strongest match), Combinational (merging two unrelated domains, like the iPhone), and Transformational (breaking the rules everyone thought were fixed, rarest and still an open question for AI).

What four things make up a person's "creative edge" that AI can't do for you?

Discernment, intuition, emotional depth, and cultural nuance.

How does the Forced Connections exercise work?

Name your challenge, add a few genuinely distant domains, ask for a handful of real insights from each domain capped at about five sentences per domain, then pick the collision worth pursuing.

Why did Sydney say a personal archive matters as "ground zero"?

Because AI is sycophantic and will back your worst ideas, so you need an external source you trust, such as your own photos, sketchbooks, writing, values, and saved references, to check AI output against.

What are the six parts of Ekta Mody Vakil's System Nucleus?

Goal, Your Role, AI's Role, Voice/Style, Decision Logic, and Sources.

Where in a workflow chain (brief, research, AI exploration, concept directions, design development, presentation) did Ekta say the one review checkpoint should go, and why?

Right after AI exploration, because that is where you decide which ideas move forward, reviewing direction, prioritization, quality, and alignment there catches the highest-risk decision instead of reviewing every draft, prompt, or revision downstream.

What is Ekta's four-question file-naming formula?

A good file name answers When, Project, Type, and Status, following the formula [Date]_[Project]_[Type]_[Status].

Why did Ekta argue that "AI doesn't fix broken workflows, it exposes them"?

Because a clear system scales cleanly while a messy one just gets amplified faster; the failures she named, like context lost between steps, feedback trapped in chats, and decisions never captured, mean the workflow itself is what breaks, not the AI.

What is the funnel Wouter Oomen traced from the raw internet to a generated image?

Common Crawl (the scraped past internet) feeds CLIP, which scores and filters caption relevance, producing LAION-5B (what survives the filter), which trains Stable Diffusion, whose output flows back into the future internet that becomes tomorrow's scrape.

What two myths about generative AI did Wouter Oomen take apart?

That gen-AI is trained on the entire internet, and that gen-AI is a mirror of society; both are wrong in specific, provable ways traceable to particular human choices.

How does Wouter's prompt-mapping method work?

Take one big concept, split it into 16 to 25 flat, specific prompts written as plain noun phrases with no editorializing adjectives, generate a grid with a raw image model, then read the grid for what it reveals.

Why does Stable Diffusion render Syria as rubble and destruction but Ukraine as churches and public squares?

Stable Diffusion's training data has a 2022 cutoff; Syria was already at war before that cutoff so war imagery dominates its data, while Ukraine's war began after the cutoff so the data still reflects peacetime scenes, since news coverage is the dataset.

What are the three canvases in Angella Tape's Strategic Foresight framework, and in what order do they run?

Category Creation first, Human Truth second, and Brand Positioning last. She insisted on starting with the category, never the product.

In Angella's working doctrine, what does the AI do and what does the strategist keep?

The AI does the data journey, the research legwork. The strategist keeps the judgment: style, experience, and the call on what matters.

What is the gap-hunting move she taught for getting past a complete-looking research summary?

Ask the model what does not exist and what data is missing. The idea usually lives in the gap, not in the summary.

Why does she phrase every instruction as what TO do instead of what NOT to do?

Because negatives confuse the model. Naming the wanted behavior works; banning the unwanted one keeps failing.

What are the eight elements of Antonio Fragoso's campaign system?

Concept and audience, voice, tension and emotion, territories, guardrails, and outputs.

What six spokes surround "concept" in Antonio's diagram of what makes copy great?

Brand voice, emotion, human tension, culture, surprise, and memorability.

What are the three judge prompts Antonio taught for evaluating a longlist of ideas?

First, kill 70% of the ideas and improve the remaining 30%, explaining why. Second, judge them like the audience would and find the gaps. Third, ask whether any of them will stand out against ALL competing content, and explain why.

Why does Antonio say "AI doesn't need better prompts, it needs better context"?

Language models remix what they were trained on and, left alone, produce the average of everything, so the unpredictable part of great copy, the emotion, human tension, and lived culture, has to come from the human work built and fed in before asking for the writing.

What do the four letters in Jaeyoung Lee's R.O.C.K framework stand for?

Representation (who the persona is), Objectives (what you want the persona to help you achieve), Core Value and Belief (what shapes the persona's world), and Key Tone and Knowledge (how the persona expresses itself).

What was the clearest example Jaeyoung used to show that a synthetic persona can be plausible but not valid?

The comparison between a real grandmother, who had no mobile phone and constraints shaped by lived reality, and a synthetic grandmother, who confidently suggested budgeting and grocery apps: technically fluent but behaviorally off.

How does the decision-logic layer that Jaeyoung added on top of R.O.C.K actually work?

It adds a trigger (when the persona acts), a first reaction (gut check), three heuristics written as if/then rules, a friction point, a default behavior, and what would change the persona's mind, so the same persona can be tuned differently (for example Martin optimized for minimum effort versus optimized for independence).

Why does Jaeyoung insist that a persona's validation never stops?

Because a persona built once and left alone drifts from the population it represents, so it needs to be iterated against real humans on an ongoing basis and asked to cite the source of any claim it makes.

What is the mediocrity loop Dom Heinrich warned the class about?

Prompting AI from your own default thinking pattern just feeds the AI back your own pattern in a loop, so the output ends up rhyming with your own habitual thinking instead of pushing past it.

What are the six Think Print archetypes Dom introduced?

Centrist (focus and control), Patternist (system and repetition), Precisionist (precision and order), Imagineer (visionary, inspiration), Interpreter (meaning and influence), and Activator (action and execution).

How does the Infinite Design process actually run once you know your Think Print stretch?

Step one is priming the machine with context it doesn't have (goal, audience, situation, known data, blind spots, what not to do) on the prompt design canvas; step two picks the thinking method (your stretch archetype as default); step three picks the generation tool per stage; step four runs about three design loops, challenging the machine's output each time before moving on.

Why does Dom say naming your default thinking archetype matters, rather than just picking a favorite?

The default is not a flaw to fix, it's a shortcut the brain built to save energy; naming it lets you notice when you're about to feed AI your own pattern back to yourself, so you can prompt in your stretch instead and get somewhere your default wouldn't naturally take you.

What are the three altitudes of Katrin Zimmermann's innovation ladder, illustrated using a taxi?

Classic design (feature and quality focus, like the taxi's yellow color), design thinking (human needs and experience focus, like Uber redesigning the inconvenience of hailing a cab), and system design/computational design (systems and data focus, like redesigning a city center for livability, growth, transportation, and climate at once).

What are the four levers of systems thinking that Katrin named?

Feedback loops (action and effect fed back into the next action), stocks and flows/balancing loops (what comes in versus what goes out), leverage points (not every lever moves the system the same amount, with mindset, culture, and trust as the strongest), and causal loops/the flywheel (a virtuous cycle once it's spinning).

In the out-think-your-own-tool exercise, what two steps did students actually perform?

First, ask a tool what happens long-term as a self-optimizing system (like Spotify, TikTok, Instagram, or Maps) keeps optimizing itself; second, and the harder part, find the loop the AI's answer left out, such as burnout, regulation, filter bubbles, or saturation, since the first pass usually only gives the virtuous flywheel.

Why does Katrin say the first-order effect of a change is rarely the one that costs the most?

Because multi-order effects, the second, third, and further-order chain reactions that follow an initial change, often outweigh the obvious first effect; her examples include the cell phone (built for communication, producing isolation) and Airbnb (more short-term rentals, depleting long-term housing).

What are the three criteria Dom and Tony use to judge the capstone work?

Scientific application of what was actually learned in the course, the outcome showing visible human thought (a generic idea reads as low effort no matter how much AI produced it), and the human-AI-human story, meaning how AI was integrated into working with other people.

What is the capstone brief, and what research signal sparked it?

The brief is 'Holiday Icon Reset,' real Coca-Cola material Dom adapted, asking students to invent a new iconic holiday element that becomes a memory structure, participation system, and commercial trigger for the next decade; it was sparked by a research signal that Santa no longer lands as cool with Gen Z.

What are Tony's four provenance questions to run on anything before shipping it?

How it was made, whether you can use it, whether you can prove you made it, and whether you can reuse it later.

Why does Tony treat it as reassuring rather than alarming that OpenAI and Anthropic each published their own AI's bad behavior?

Because publishing the failures openly, such as a model deceiving evaluators or an Anthropic test where a model threatened blackmail to avoid shutdown, is 'not a cover-up, that's speaking out loud,' meaning the labs are showing the system of disclosure and alignment testing is working, not hiding the problem.

What are the eight fields on the custom GPT configure screen that Dom walked through?

Name, Description, Instructions, Conversation starters, Knowledge, Model, Capabilities, and Actions.

What is the key difference Dom drew between a Project and a custom GPT?

A Project is static, everything inside it inherits the same instructions, tone, and behavior, good for one campaign that should always sound the same; a custom GPT is portable, it can be pulled into any chat with one click, or several can be added at once to argue with each other.

What are the four steps of vibe coding that Dom demonstrated?

Sketch it on paper, photograph it, ask for HTML, CSS, and JavaScript, then refine with detail and interactions.

Why does Dom describe his sketch to AI as if explaining it to someone who is blind?

Because build accuracy follows description accuracy, not the sketch alone; explaining every detail in words the way you would to someone who can't see the sketch forces precision that stops the machine from guessing.

What are the five stages of Angella Tapé's brief-mining process using the Creative Collaborator GPT?

X-ray the brief (find the tension it avoids stating), escape the category (break the conventions), spar with your rebel strategist (score against the creative ladder), bring it all together (state the idea), and build a unique brand voice (a bonus stage not run in class).

What design constraints did Angella build into the Creative Collaborator GPT to keep the user in control of the pace?

A maximum of two questions per section, a required approval gate before advancing to the next stage, outputs pre-formatted as tables, and questions sourced from interviews with working creative directors.

How does the creative ladder scoring in stage 3 actually work?

Every idea is scored 1 to 10 against the ladder, with the target zone being 8, 9, or 10 (contagious, cultural phenomenon, legendary); anything scoring below 8 gets killed, and no AI is involved in this elevation round, only humans discussing, protecting, combining, and pushing further.

Why did Angella say that every group's brief-mining output converging on the same ideas was itself a finding, not just a coincidence?

Because same source material, same model, and same guardrails produce correlated output; the convergence demonstrated live the course's multi-order lesson that shared tools correlate everyone's results, which is why the room pushed for human forced connections instead.

What are the five acts of Sharon Panelo's business story structure, borrowed from Shakespeare?

What is (exposition, the world today stated plainly), the tension/the enemy (rising action, the thing to push against), the truth/the insight (climax, the insight that flips the tension), what could be (falling action, the vision of the promised land), and the call to action (resolution, the urgency that turns the audience into participants).

What are the three things Sharon said a story must measurably change for it to count as a story?

What people THINK, what they FEEL, or what they DO; if a story moves none of them, it is not a story yet.

How does the Future Headlines exercise actually work?

Jump forward to December 2030 and write the culture, business, and human headlines the idea would earn, then rewrite the best headline in multiple publication voices (such as The New York Times, NBC, NPR, and ESPN), and finally ask what would have to become true for those headlines to exist.

Why does Sharon insist villain mode should only be used at the end of the process, never at the start?

Because villain mode is meant to anticipate objections and strengthen a mostly finished idea by attacking it as a cynical cultural critic and writing the failure headline; used at the start it would only kill ideas before they had a chance to develop, so the follow-up move is always asking what you'd change so the criticism is no longer true.