The reading
The essays the argument was made in first, and the outside work the book leans on. Every row carries a line saying why it is worth the click.
Five pieces from 2026, in the order they ran. This is the argument before it had chapters, so some of it is looser than the book and one or two are wrong in ways the book goes on to correct.
- Taste Is What You Don't Ship AI made design output free. The job flipped. Now the work is killing 39 directions on purpose so the right one can ship.
- When Taste Becomes a Tax Taste matters. So does scale. The chef who refuses to serve the chicken sandwich is right about taste and wrong about everything else.
- The Taste AI Can't Train On Suhail Gupta is right that individual designer taste won't save you from AI. He missed the taste AI can't ingest because it was never published, the one your team argues out in a closed room every Friday.
- Define Your Conversational Taste Before Your AI Does It For You A four-hour workshop that ends with Claude writing in your voice, and a one-page spec that runs your humans and your AI on the same playbook.
- Three Designers and a Thousand Interns What a design team looks like when the model does the making, and the one job that's left for you.
Fifteen links in three groups: what other people have already standardised, the case for writing taste down at all, and what the research actually found. These are the ones I keep going back to.
1: The taste-file standards
Other people have already argued about what belongs in a file like this. Read one before you write yours, and steal the vocabulary.
- TASTES.md The standard the book borrows its structure from: hard rejections, thresholds, reference anchors, and the tensions you refuse to resolve.
- PitayaK/taste.md A personal taste file the agent reads at the start of a session and updates as it watches you work. The version of the idea that maintains itself.
- Taste Skill An installable skill aimed squarely at the flat default look. Worth it if you would rather correct a maintained baseline than start from a blank page.
- The Awesome DESIGN.md collection Fifty-eight real products, from Linear to Stripe to Airbnb, with their visual systems written as rules an agent can read. The mood board done right.
- Installing skill files in AI coding agents The plumbing: where the file has to sit for each agent to find it, and why an aspirational rule has to become a measurable one. The 4px example in chapter 4 is theirs.
- Ten AI skills for real design taste A survey of what people are actually installing. Useful mostly as a map of which problems the community has decided are worth a skill.
2: The argument
Why the model produces the flat thing, and why the fix is writing down what an experienced designer would otherwise infer.
- Why AI design looks generic Chapter 2 in someone else's words: one prompt asked to do taste, exploration, and spec at once, and a model that settles the conflict by heading for the middle.
- Five shifts redefining design systems in the AI era The whole book in one line. Anything an experienced designer would normally infer has to be made explicit: the context, the constraints, the examples, and the reason behind each.
- I am just a text file Ruben Hassid's hundred-question voice extraction, built for writing. The design taste interview on this site is an adaptation of it, so read the original to see what it does.
3: The evidence
The studies underneath the claims. Two of them are the reason the book refuses to let a model grade its own work.
- You never stop cultivating taste Loredana Crisan's test: not whether everyone likes the work, but whether you can say what is not working and why. That is the test a taste file has to pass in writing.
- Aesthetic taste and its limits UC Berkeley on the tacit ceiling: when the only channel is a prompt, designers cannot say what they mean about space and material. Why chapter 5 leans on anchors, not adjectives.
- TASTE: a designer-annotated preference dataset Ten designers ranked AI-generated work on nine criteria. They agreed better than chance and worse than you would hope, and no automated judge cleared 0.55 agreement with them.
- The double-edged roles of generative AI Generative AI helped during ideation and hurt during convergence, with professionals doing measurable extra rework. The cost of accepting the first plausible output, priced.
- Algorithmic environments and aesthetic autonomy Links recommendation-shaped exposure to homogenised taste among design students. Survey-based, so take it as a reason to widen your references rather than as proof.
- Does mode of evaluation impact decision-making? Judging two options side by side changes both the choice and the reasons given for it. Why the book asks for pairwise comparison instead of scoring everything out of ten.
Add unlimited notes to any activity
Comments & Discussion
Add a Comment
Recent Comments (3)
This workshop was incredibly effective for our remote team! We adapted it slightly for a virtual setting and it worked wonderfully. The key was breaking into smaller breakout rooms.
Great resource! One tip: prepare all materials the day before to avoid any last-minute rushes.
Used this for our quarterly planning session. The structured approach really helped us stay on track!