This is the most valuable article I’ve come across recently — sharing it with everyone.

A discussion from Wharton on how to develop software in the AI era.

They built an AI studio (Generative AI Studio) to observe how students use AI to create projects.

They observed that two typical collaboration patterns limit students’ creativity:

1. Students are always ready to act, eager to turn ideas into reality — but this prematurely locks their energy into execution rather than invention.
2. When students brainstorm together, groupthink quickly emerges — mediocre ideas get echoed and amplified, while truly original ideas become scarcer.

At the same time, students’ attitudes toward AI are equally typical:

1. Students treat AI like a vending machine, expecting a simple prompt to yield a perfect finished product. They don’t manage AI’s stochastic nature and lack patience for the ongoing dialogue, added context, and incremental course-correction needed to get good results.
2. Students treat AI as just a code-writing tool. One student wanted to build a fitness app and immediately dove into designing databases and saving workout data — without thinking about how to use AI to create experiences ordinary apps can’t offer.

In response, the Wharton faculty proposed a solution.

Judgement, taste, creativity, and critical thinking remain the most valuable traits in the AI era. Starting from this premise, they designed two practices: Crit (critique) and Charrette (a one-day time-boxed creation exercise).

Crit — Developing Judgement

Think of it as peer review of creative work. There are two formats:

1. 1-on-1 crit: An instructor gives direct feedback — what needs strengthening, what’s working well, whether a change of direction is needed, whether to go back to the drawing board, etc.
2. Group crit: Facilitated by an instructor, peers review each other’s work.

Regardless of format, the opening is the same: students present their work first, then solicit feedback from the room. During the presentation, they need to address five things:

1. My intention — what I set out to do
2. My process — how I made it
3. What I discovered along the way
4. What new questions emerged
5. What challenges remain

To keep feedback quality high and prevent peer review from becoming a kangaroo court, instructors require all comments to begin with one of three phrases:

1. “I notice…” — Slow down, look carefully, describe what you want to discuss rather than jumping to conclusions.
2. “I wonder…?” — Open a dialogue with a question rather than closing one with a verdict.
3. “What if…?” — Present alternative directions as possibilities, helping the other person advance their own goals or discover unexpected opportunities.

Here’s a real example:

A student wanted to build a fitness app. His approach was conventional: a questionnaire + asking about gym equipment + AI generates a workout plan. During a Crit, someone pressed: “Could AI create experiences ordinary apps can’t?” This question directly shattered the assumption — no equipment database was needed. The user simply takes a photo of their surroundings, and the AI understands the context and generates a truly personalized workout plan.

This example is very compelling. I frequent hotel gyms, and if an app like this existed, I would genuinely use it.

Expanding outward, we could adopt this kind of review process within teams or companies as well — highly referenceable.

Charrette — Developing a Creative Process

The Charrette process has six steps:

1. Set the problem
2. Individually, using pen and paper, diverge on as many rough ideas as possible
3. As a class, curate and evaluate the ideas, then decide which products to develop
4. Teams of 2–3 do planning only — no execution whatsoever
5. Using AI, build a prototype within a two-hour time limit
6. Class-wide presentation and critique

The entire process is limited to 8 hours, with the build portion capped at just two hours.

This is my favorite part of the entire article. If you’ve done any development work, you’ll easily recognize how much time you sink into execution — sometimes for days on end without stopping. But when someone asks if you have a working prototype to try, you often realize you have nothing to show.

The point of the Charrette is to make students realize that in the AI era, planning work matters more than execution work. AI can handle execution for you — you shouldn’t place yourself in the executor’s seat anymore. But AI can’t replace your creativity and critical thinking. So you should give yourself more time to think about the problem and the solution, rather than rushing to start building.

A great article, well worth sharing. Original link:

https://gail.wharton.upenn.edu/research-and-insights/enduring-methods-emerging-medium/…

Enduring Methods for an Emerging Medium: The Generative AI Studio

Source: https://gail.wharton.upenn.edu/research-and-insights/enduring-methods-emerging-medium/

Upskilling efforts are teaching workers to use generative AI tools, but the rise of agentic work is driving a call for enduring skills: judgement, taste, creativity, and critical thinking. We have not seen a pedagogical approach designed to develop these skills in a way that could realize generative AI’s full potential and the economic and human value that could come with it. So, we built one, and it begins with changing how we define generative AI itself.

Generative AI is more than a tool for productivity and efficiency. It can produce stories, images, music, and interactive experiences that stir human emotion. It is stochastic and, like paint or film, what it produces is heavily shaped by the person who directs it. These are the qualities of a creative medium, not a tool. So, we treated it as one, applying the time-tested methodologies of art and design education to teach students to work with it.

In the Generative AI Studio, students pursue individually driven projects with no right answer, only the potential to serve a need or create a meaningful human experience. Two practices were core to the educational experience: the critique, which develops discernment, and the charrette, which develops a productive creative process. Together they enabled students to create an unexpected array of AI-driven experiences and surfaced insights that hold lessons well beyond the classroom.

Why The Arts?

Art and design education leverages repeated cycles of creation, critique, and revision to build the very skills that working with generative AI demands: imagination, technique, process, taste, judgement, and critical thinking.

The Crit: Developing Discernment

Crits center on the assessment and discussion of creative output. Their purpose is to help the student clarify their own goals and identify ways to build on what is already strong within the project. Through repeated crits, students gain multiple perspectives on their work, build a shared critical vocabulary, and develop the ability to discern what success looks like while shaping their own taste. (How we structure and facilitate crits, including the norms that keep them psychologically safe, is covered on our [Pedagogy page (https://gail.wharton.upenn.edu/gen-ai-studio/the-generative-ai-studio-pedagogy/)].)

The Charrette: Developing Process

Discernment is only one half of the picture. The other is process: the question of how we approach our work with generative AI.

We observed two patterns in how our students worked that were clearly limiting their creativity.

• Our students were poised for action and eager to make ideas real. This was often a strength, except when it prematurely focused their energy on execution, rather than invention.
• When students brainstormed all-together, as often happens in a class or a meeting, groupthink emerged quickly, reinforcing weak ideas, and producing fewer truly original ones.

The charrette, a single-day exercise in developing novel solutions to a specific problem within a fixed window of time, addressed both issues.

Anatomy of a One-Day Charrette: From Idea Generation to Project Presentation in 8 Hours

1. Setting the Stage The problem is framed, new tools are introduced, and the criteria for success are made clear. Whole Group

2. Divergence Individuals generate as many rough ideas as possible, none prejudged. The work stays deliberately analog: paper, post-its, and whiteboards. Each Student Alone

3. Sense Making A mini-crit of the ideas. Patterns surface, and instructors review all of the analog work to select the ones worth developing further. Whole Group

4. Convergence Small teams are formed, each takes one curated idea. This stage is planning only — no code yet. Teams of 2 or 3

5. Rapid Prototyping Up to two hours to make a working prototype. The thinking is already advanced; here the build tools come into use and code serves the idea. Teams of 2 or 3

6. Assessment A closing crit where prototypes are presented. Seeing one problem solved many ways makes visible the range of possibilities. Whole Group

Afterward, the work was more varied and inventive, students’ expectations for themselves changed, and they left with a concrete ideation process they could carry into other settings.

What Does Success Look Like?

Core to the crit process was a set of criteria that we used to assess our work.

• Intrinsic AI: A clear, essential role for generative AI, without which the project would lose its distinct value.
• Aesthetic Elements: The effective use of aesthetic elements, including image, color, typography, composition, time, motion, and sound.
• User Interface: A logical user experience that serves the project goals and maximizes the distinct value created by generative AI.
• IT Architecture: A logical and efficient digital infrastructure.
• Data Security & User Safety: Security and safety regarding data, privacy, and unintended consequences on others.
• Ethics: Ethics regarding the use of images, music, video, and text produced by generative AI, or other creators.

One student project revealed how the crit can spark innovation. They were building a fitness app for frequent travelers who need to adapt their exercise routine to their current environment. The initial iteration was designed like most apps: it had simple click-through, multiple choice surveys that enabled the system to pull a sub-set of options from established data sets to provide the user with a “custom” set of solutions. In pursuit of this, the student focused first on acquiring access to gym equipment databases rather than developing the user experience. AI was being treated as a tool to rapidly write the code that would drive the app but not part of what made the app unique.

In the crit, a new question emerged that challenged this approach: what if generative AI could enable experiences that ordinary app logic could not? Could it create truly personalized solutions? Through our probing of preconceived notions, like the need to have a database of equipment from which a workout could be planned, and asking open-ended questions about what might be possible, the survey gave way to a live exchange between the user and the AI. A simple photo of the user’s immediate environment was all that was needed for the system to create a truly adaptable workout plan. The AI was no longer just a tool to build the app; it became a medium through which the traveler’s unique conditions could be understood and turned into a workout.

Embracing Process

Most students initially approached generative AI expecting that a simple input would produce a polished result, like putting a dollar into a vending machine and expecting it to deliver a bag of chips. As the crits helped students become more discerning regarding the quality of AI’s output, they realized that they needed to learn to manage generative AI’s stochastic nature and jagged edge to get better results. So, we taught them to treat prompting as a management process that centers on good communication and collaboration: explaining to the model what they were trying to accomplish, sharing the questions they were wrestling with, and asking the model to help shape and improve its own instructions. Over time, students began to approach generative AI more like a member of their staff, providing direction, context, and ongoing course correction. As they did, the quality of their output rapidly and significantly improved.

Invest in Iteration

We also had to pull students away from building too early, and teach them to invest in the iterative process of idea development. When they focused on the code first, which was their first instinct, their core concepts remained underdeveloped while the architecture and code base became solidified, triggering the commitment bias that makes it challenging to reconsider the original idea. When students stopped treating generative AI mainly as a coding system and instead approached it first as a medium for shaping human experience, the work that resulted was more distinct and compelling, and the build that followed was more efficient and effective.

By the end of the semester, students had advanced both the creativity of their ideas, and the forms that their projects took, moving well beyond the chat interface. Their projects involved speech, vision, robotics, XR environments, and dynamic interfaces. This included a voice-driven system for shaping three-dimensional objects in a virtual environment, a city guide that curated experiences for travellers that were informed by residents’ own stories, and a system that renders current events into age-appropriate, educationally sound, politically neutral comics for children.

Key Takeaways

The crit is a path to discernment.

Structured critique helps students clarify their goals, build a shared critical vocabulary, and learn to weigh feedback, while shaping their taste and definition of what success looks like.

The charrette develops process.

A fixed-time cycle of divergence and convergence gives students a concrete ideation process they can carry into other settings, and counters the pull toward premature building.

Management skills matter.

Learning to work with generative AI means learning to manage it: providing direction, context, and ongoing course correction, and developing the instructions themselves as an early product of the process.

Building too early is a trap.

While generative AI makes coding almost instantaneous, and prototyping faster than ever, hardening the technical structure around a weak or untested idea stalls meaningful innovation.

Enduring methods can unlock emerging technologies.

Art and design education has leveraged studio-based instruction to cultivate enduring skills for generations. When we treat generative AI as a creative medium rather than a productivity tool, those methods apply directly, preparing students not just for this technology, but for whatever comes after it.