Feeding References to AI
Describing style by mouth alone is inefficient. When to use reference images, style descriptions, or design variables — plus templates — so AI ships from your reference library
THE QUESTION THIS PAGE ANSWERS
ANSWER FIRSTWhat is the key idea behind “Feeding References to AI”?
Describing style by mouth alone is inefficient. When to use reference images, style descriptions, or design variables — plus templates — so AI ships from your reference library
Turn taste into a behavior the product can repeat. The useful outcome is not a nice opinion. It is a visible rule, a small example, and a way to tell when the experience falls below the bar.
Capture one before-and-after example that shows the quality bar without extra explanation.
Polish that improves the surface while leaving the user's uncertainty untouched.
"A bit warmer, a bit more premium"—say it ten times, nine warp inside the model. Adjectives take two translations: you compress feeling into words, the model expands words into pictures; both ends lose fidelity. Feed references straight to AI—least loss, one shot.
Three postures for feeding references—each covers a stretch. Pick from the table:
| Method | What you give | Best for | Weak spot |
|---|---|---|---|
| Reference image | 1–3 reference images | Overall vibe, composition, style transfer | Easy to copy layout too |
| Style description | A spec paragraph: density, primary, radius | Styles you can state as rules | Vibe you can't spell out stays unwritten |
| Design variables | Token list: primary / radius / type / spacing | Having AI write UI code | Locks style, not layout or copy |
You don't invent style descriptions from scratch—there's a method. About Face 4, Chapter 17, records Cooper's practice: experience attributes. Before designing, pick 3–5 adjectives with the client that describe product vibe and brand promise—"clean, restrained, trustworthy." Once set, those words referee every visual call: when unsure, ask whether the words would approve.
Adjectives are allowed to fight. The book says "safe" and "flexible" can both sit at the table—keep that tension: where two words clash is exactly what early style drafts should answer first.
Moved to feeding AI, it fits flush: pin the vibe with adjectives, then translate each into a concrete spec. Adjectives alone, AI reads the average; finish the translation step and the style description is done. The workbench below does that translation.
Brief: home for a budgeting app, vibe locked as "clean, restrained, trustworthy." Version A sends the three words as-is; B runs the workbench first, translates word by word into specs, then sends. Tap the one you think is better.
Each feed method gets a ready prompt. Pick a method, copy, swap the placeholders, send. For style description, drop the workbench output straight into the body.
The third method deserves its own line. Tear a reference into design tokens—primary, radius, type, spacing each become a variable—and what you hand over is more than a reference: it's an interface standard.
Why are standards valuable? About Face 4, Chapter 17, cites Nielsen: a unified interface standard helps users learn faster and err less, because experience in one place predicts behavior elsewhere; for the team, ready decisions skip round after round of debate. The same ledger holds for AI: once the token list is in, every generation lands on the same standard—ten revisions won't drift, and you have a yardstick at review.
The same chapter says the hard part first: follow the standard unless you have a strong alternative. Breaking is allowed—reasons must be hard. That rule fits you and AI alike. One last duel settles this lesson's ledger on the spot.
Brief: brand page for a bakery studio. A's input is only "a bit warm, a bit premium"; B brings a reference image and a variable list. Tap the better version.
Adjectives warp; references don't. Stuff images, spec text, and variable lists straight into the chat—least loss.
Three feeds, three stretches: Reference images cover vibe, style descriptions cover rules, design variables cover code. Combine if you want—just know who owns what.
Style description has a method: Pick 3 adjectives to pin the vibe (experience attributes), then translate word by word into specs. Words referee; specs execute.
A token list sets the standard for AI: Value is predictability—every generation lands on the same variables, drafts don't drift, review has a yardstick.
Mediocre output? Check the input first. Adjectives alone land on the average; swap in references and variables and the same model changes face at once.
Source: Original to Xiaoshan Academy's Taste Engineering series; some design principles adapted from About Face 4, Chapter 17 (Alan Cooper et al.); experience attributes and standards discussion from the same chapter.
Turn the feeling in “Admit it first: adjectives never reach the AI” into a judgment
“"A bit warmer, a bit more premium"—say it ten times, nine warp inside the model.” points out that AI has lowered the bar for making something usable. The skill readers need is noticing what is wrong and turning that feeling into an actionable requirement.
Watch the user's next action, not just the surface
Turn “Three postures for feeding references—each covers a stretch.” into observable questions: does the user know what happened, what to do next, and how to recover from an empty or failed state? Does the hierarchy make the important information visible first?
Pretty is not the same as usable
Apply “Mediocre output?” to a second screen or flow. Record one moment of hesitation and the user action after the change; observable behavior is stronger evidence than polish alone.
From “Admit it first: adjectives never reach the AI” to “Try it · Scenario match: three needs, one feed each”
“Admit it first: adjectives never reach the AI” grounds the problem in “"A bit warmer, a bit more premium"—say it ten times, nine warp inside the model. Adjectives take two translations: you compress feeling into words, the model expands words into pictures; both ends lose fidelity…”. “Try it · Scenario match: three needs, one feed each” then moves it toward “Have AI redraw my event poster and keep brand vibe Single choice A Reference image: attach past posters so it learns the vibe B Style description: write a "premium and grand" blurb C Design variables: hand over…”. Together, they show that the lesson is not just a conclusion to remember, but a claim with conditions.
Carry the judgment into the next situation
For experience work, turn abstract impressions into user actions: did the person understand the state, find the next step, recover from an error, and want to continue?
- “Admit it first: adjectives never reach the AI”: "A bit warmer, a bit more premium"—say it ten times, nine warp inside the model. Adjectives take two translations: you compress feeling into words, the model expands words into pictures; both ends lose fidelity…
- “Try it · Scenario match: three needs, one feed each”: Have AI redraw my event poster and keep brand vibe Single choice A Reference image: attach past posters so it learns the vibe B Style description: write a "premium and grand" blurb C Design variables: hand over…
- “The closing point”: The third method deserves its own line. Tear a reference into design tokens—primary, radius, type, spacing each become a variable—and what you hand over is more than a reference: it's an interface standard
The final “The closing point” brings the discussion to “The third method deserves its own line. Tear a reference into design tokens—primary, radius, type, spacing each become a variable—and what you hand over is more than a reference: it's an interface standard ”. The useful thing to carry forward is knowing which judgments must be revisited when input, scale, or risk changes.
I turned one judgment from this article into a small experiment I could run today. Knowing what to observe next is more useful than simply remembering the conclusion.
After reading this, I first looked for the conditions behind the idea instead of copying the method into a project. That order made the later trade-offs much clearer.
When this judgment reaches real work, which constraint should be added first? I am curious which step matters most between reading and the first practical attempt.
No discussion on this article yet.