The CPA Rubric
Cognitive-Process Assessment
Grading the thinking behind AI-assisted design, not just the artifact
The eight dimensions
Each one carries its own evidence trail. The tool column is what my own studio sequence uses, and it is the first thing you should replace.
| Dimension | What it measures | Evidence | Tools |
|---|---|---|---|
| Prompt Architecture | Strategic intent, vocabulary depth, and the context framed behind AI instructions. | Raw prompt logs, systematic variable testing, multi-turn iteration history. | Google Stitch, Gemini |
| Orchestration Logic | How the student chains multiple AI tools into an end-to-end design workflow. | Workflow maps, tool-handoff documentation, architectural decision rationales. | Gemini Notebook |
| Strategic & Content Development | Critical thinking in making and justifying content decisions, feature scope, content priorities, and value proposition, grounded in research findings rather than default or generic choices. | Feature-scope prioritization frameworks, content and IA decision logs, value-proposition rationale. | Gemini Notebook, Gemini |
| Critical Curation | The student's ability to evaluate, stress-test, and refine raw AI output against constraints, naming each catch as an Attribution Error, a Logical Error, or a Representational Error. | Revision trails, error-tracking sheets, source-verification and fact-check logs, written critiques of AI-generated content. | Gemini |
| Creative Art Direction | Taste, aesthetic consistency, and conceptual choices that override default AI biases. | Annotated style matrices, mood boards, mood-to-asset translation logs. | Google Stitch |
| Domain Specific Knowledge Application | Visual communication craft, typography, hierarchy, color, and compositional judgment, applied well enough to turn AI-assisted output into original, intentional design work. | Typography and hierarchy exploration notes, color and composition rationale, design system documentation, before and after craft comparisons. | Google Stitch, Figma |
| Reflective Reasoning | The student's ability to explain why they made specific curatorial choices, as a legible reasoning chain where each step is independently defensible, not just the final conclusion. | Timestamped process notes, file annotations with rationale, spoken walkthroughs in critique. | Gemini Notebook, Gemini Gems |
| Direct Creative Contribution | Human authorship. What the student sketches, develops into original artifacts or moodboards, writes, or fundamentally transforms. | Original source files, before and after authoring comparisons showing hand-finished or transformed work. | Gemini Notebook, Google Stitch, Figma |
Three kinds of error
Two are Google's own language, from the REVEAL paper on verifying reasoning chains. The third came out of my classroom.
Attribution Error. A fact the AI made up. A claim or source that simply isn't true, and wasn't caught.
Logical Error. A step that doesn't follow. A flawed deduction, even from true premises.
Representational Error. New, and provisional. A default the model keeps returning. A demographically skewed output, repeated against an explicit instruction to do otherwise. It is not a fabricated fact and it is not bad reasoning. It is a different failure and it needs its own name.
Jacovi, A., Bitton, Y., Bohnet, B., Herzig, J., Honovich, O., Tseng, M., Collins, M., Aharoni, R., & Geva, M. (2024). A Chain-of-Thought Is as Strong as Its Weakest Link: A Benchmark for Verifiers of Reasoning Chains. Proceedings of ACL 2024. arXiv:2402.00559.
Where the assessment happens
In critique. In every class, students discuss the what and the why of their work. Design education has run on critique for a century, so CPA does not add an event to your calendar. It adds a rubric to a conversation you are already having, and that recurring conversation is what dates the record.
And it cannot be only written. A written log can be generated in seconds. A spoken one cannot, because you can stop a student and ask why they rejected the other three directions, and both of you find out in real time whether they know. Live explanation is also where the thinking happens, not only where you check it.
Thalia Merino, FIT. She had never seen this rubric.
Make it yours
This rubric was built for a UX and UI studio sequence. It is not a standard and it is not finished.
- Reweight the dimensions your course actually cares about.
- Drop the ones it doesn't. Eight is what my course needed, not a magic number.
- It does not have to stay in design. The structure is: name the decision, show the evidence, explain it out loud.
- Change the tool column entirely. The dimensions are about judgment, not about which products a student used.
You do not need my permission and you do not need to tell me. If you do test it, I would like to hear what broke.
Where this came from
CPA is Chapter 2 of Empowering Design Education with AI: Curriculum Innovation in the Age of Creative Intelligence, by Christie Shin and C.J. Yeh. The rubric and the research are open source. The book is a separate purchase.
Built during the Google Higher Ed Faculty AI Fellowship, Inaugural North America Cohort 2026, and tested in CT302, Digital Product Design One, at the Fashion Institute of Technology. The empathy interviews behind it and the Institute materials are on this site.