Design by Selection

Sources#

Summary#

The working practice Nate Parrott describes after a year of daily use of Claude Design — the tool he built (case-study, first-hand, unmeasured). Its shape is a barbell: heavy human effort at the front (deciding what you actually want, usually away from the computer) and at the back (hand-editing the last mile), with the expensive middle — producing candidate artifacts — generated ten at a time and selected from rather than authored.

"Ask for ten options, then remix. Most of them won't be good, and that's fine; one or two will be. Then say, 'I like option B and a little of option D. Give me five riffs that smoosh those together.'"

The designer's scarce resource stops being production and becomes discrimination plus specification: knowing what you want precisely enough to ask for it, and recognizing it when it appears among nine near-misses.

The barbell#

Front end — do the thinking before you prompt. "I spend a lot of time writing prompts before I design." The methods are deliberately off-keyboard: dictating into the tool's voice button, typing into the phone's Notes app from the couch, recording a voice note on a walk and pasting the transcript later. The rule generalizes past the channel: figure out what you want while you're away from the computer, so the model can execute your exact vision when you sit down. Sitting at the tool and thinking through it wastes the fast part on the slow work.

Middle — generate wide, then remix. Ten options, expect one or two to land, then ask for riffs that combine them. Note what this assumes: that judging ten candidates is cheaper than authoring one. That holds for visual work and is the whole reason the move is rational.

Back end — make the last mile manual. Use direct manipulation (rearrange, delete, edit text, resize, recolor) for final touches instead of prompting for them, for two separate reasons: "direct edits use no tokens, and small calls like sizing and alignment are better eyeballed anyway." The second reason is the durable one — some decisions are faster to make than to describe, and no model improvement changes that. This is a clean counterexample to the assumption that a better model absorbs every step of the loop: the last mile is not a capability gap, it is a bandwidth mismatch between pointing and prose.

Two supporting techniques serve the same bandwidth argument: sketch what you can't describe (draw the layout on paper, photograph it, upload) and point and talk (click the element, speak into a comment rather than writing a paragraph identifying which element you meant).

The default aesthetic is a real failure mode#

"Left undirected, Claude picks one of its favorite aesthetics. You'd probably recognize them."

Stated flatly, by the person who built the tool, with a link to press coverage of the house style. This is direct first-party evidence for reason 3 in Why AI Lags at Design — design rewards novelty where code rewards known patterns, so regression to the mean is a feature for code and a failure for design. Ambrosino's version was an argument about why models lag; Parrott's is the same effect observed as a daily nuisance requiring active countermeasures:

Specify fonts and colors explicitly.
Provide a moodboard of images for inspiration.
Have Claude brainstorm font-and-color pairings and iterate until one feels right — using the model to escape its own default rather than accepting the default.

The escape route generalizes: the aesthetic default is what you get when the specification is silent on aesthetics, so aesthetic direction is not polish applied at the end but a load-bearing input at the start.

Where the escape route stops. An independent third-party test — Evangeline driving Meta's Astryx design system against Salt & Straw (I tested Meta's "agent-ready" design system Astryx. Here's the results., case-study, n=1) — puts a boundary on it. Fully specified brand artifacts (DESIGN.md, layered token files) did defeat the default aesthetic: correct coral, correct flat-card depth, correct typography, across a full multi-screen product. But Astryx exposes deep customization for only four components, and every component outside that set — progress bars, status labels — rendered in Astryx's own default blue regardless of the specification. The collapse is therefore fixable by context where a slot exists, and untouched by context where one doesn't: the failure is a coverage boundary in the design system, not a taste ceiling in the model. Notably the model read the brand correctly throughout; it had nowhere to apply it. Two corollaries worth carrying: the silence is the danger (a brand whose colors happen to resemble the system's defaults looks like success — Astryx passed with Blue Bottle for exactly this reason), and specification can be violated even inside covered surfaces (a segmented control picked up a shadow despite an explicit "no shadows, ever" rule). See Living Design System for the full test.

Fidelity as a control knob#

"Wireframe first when fidelity doesn't matter. Asking for wireframes is much faster, and it keeps Claude focused on the higher-level structure of a design instead of the visuals."

Two things at once: a speed optimization, and a steering mechanism — withholding fidelity keeps both the model and the reviewer on structure. It is the practitioner's answer to Polish No Longer Signals Readiness: when polish is free, it stops encoding process stage, so a designer who wants structural feedback has to deliberately suppress the polish that would otherwise arrive by default and pull the conversation toward the wrong layer.

Feeding it the real context#

"Turn recurring work into a design system" — upload brand files, logos, decks, screenshots, typography specs, and Claude generates a design system so "each artifact you make afterward starts from your choices, rather than a blank slate." Connect GitHub and it fetches existing components and screens; "with a few tries it can recreate your existing designs with pretty high fidelity."

This is Living Design System arriving as a first-party product feature rather than a hand-rolled design_system.html, and it is the same fix as the aesthetic-default countermeasure one level up: replace the model's priors with your own artifacts.

Building the tool to make the artifact#

The intro animation for Claude Design was made in Claude Design — but not directly: "I'm not an animator, so I first had Claude Design build me a bespoke video editor, then used that editor to make the animation." A disposable micro-app one level deeper than the usual case: not a UI to edit a plan, but an entire authoring tool for a medium the human doesn't have skills in, generated so the human can then work by direct manipulation in a domain where they can judge but not produce. It is the barbell's back end taken to its conclusion — when the last mile needs a tool that doesn't exist, generate the tool.

Where the practice does not apply#

Parrott is explicit about the boundaries, which is unusual in a first-party post:

No image model. Claude Design "isn't built for image generation, so it's a poor fit for logo design — though that hasn't stopped people from trying." Bring the logo you have.
Not for shipping production software. "If you're shipping production software, stick with Claude Code." The split is ideation vs. production, with a round-trip in both directions, not a capability ranking.

Connections#

HTML as the New Markdown — an independent arrival at the same premise from a different Anthropic practitioner: Parrott's breakthrough was "Claude is really good with HTML… we think of HTML as the format for websites, but it's also a rich, interactive visual medium — anything you can make in a slide deck, a video file, or a PDF, you can make in a web page." Thariq Shihipar reached this for plans, Parrott for design, and neither cites the other
Why AI Lags at Design — the default-aesthetic collapse is first-party confirmation of reason 3 (design rewards novelty; models regress to the mean), observed from inside the tool built to close the gap
Polish No Longer Signals Readiness — wireframe-first is the working countermeasure: deliberately withhold fidelity so the artifact reads as early and the feedback lands on structure
Living Design System — the same portable-design-DNA move, productized: upload brand files, get a generated design system every later artifact starts from
Disposable Micro-Apps — the bespoke video editor extends the pattern from editing a plan to authoring in a medium the human lacks skills in
Prototype Over PRD — "generating 15 versions of a flow to collect feedback from colleagues" is the prototype-as-spec loop at option-set scale; the artifact circulated for buy-in is the spec
Implementation Abundance Inverts Product Work — Parrott states the thesis directly: "as models get better at building production software, the work that matters most moves earlier in the process — having good ideas, getting everyone aligned, and collecting feedback while an idea is still early"
Claude Design — the tool this practice is built around, described by the designer who built it
Nate Parrott — the practitioner
Claude Opus 5 — the vision gains that widen what the practice can do: better at reading charts, diagrams, and screenshots, which is the input side of design work

Open Questions#

Is "make the last mile manual" durable or transient? Parrott gives two reasons (tokens, and eyeballing beats describing); the token argument dies with cheaper inference, the bandwidth argument shouldn't. Which one is actually load-bearing is testable by watching whether direct-manipulation use falls as models improve.
Is the default-aesthetic collapse fixable by context (brand files, moodboards) or is it the novelty ceiling from Why AI Lags at Design reason 3 — i.e. does a fully specified design system still produce house-style output underneath the palette? #oq/source Partially answered: I tested Meta's "agent-ready" design system Astryx. Here's the results. (n=1, single practitioner) says fixable where the design system has a slot for the decision, untouched where it doesn't — the model read the brand correctly throughout and only regressed on components with no customization surface. That reframes the residue as a coverage gap rather than a novelty ceiling, but it tests one vendored system, not the underlying question of whether house style persists on fully covered surfaces.
Ten-options-then-remix assumes the human reliably recognizes the good one. Where does selection break down — does discrimination degrade when all ten candidates are competent, and is there a candidate count past which review cost exceeds authoring cost?

Sources#

How the product designer who built Claude Design uses it — Nate Parrott, Anthropic, July 2026 (case-study): the ten best practices, the default-aesthetic admission, the scope boundaries, and the bespoke-video-editor account
I tested Meta's "agent-ready" design system Astryx. Here's the results. — Evangeline, Substack, July 2026 (case-study, n=1): third-party test bounding the default-aesthetic escape route — context wins where the design system has a customization slot, defaults win where it doesn't