## The Corpus Problem[#](#the-corpus-problem)
If you ask an AI to design a UI without any guidance, you’ll often see the same patterns: dark mode, rounded corners, subtle gradients, and maybe a glow behind the main element. You get pill-shaped tags, popular fonts like Inter or SF Pro, and a standard 12px card radius. The result looks good and works well, but it blends in with everything else. This is the main issue: when AI design isn’t guided, it settles for the average. You don’t have to stick with that, but you do need to work to break out of it.
This challenge arises when utilizing the entire corpus as a source. The designer is effectively competing with an aggregated, distilled version of the internet’s collective design decisions. The standard output represents the most common, rather than the most effective, design choices. Employing a frontend-design
skill or a deny list of patterns addresses the issue superficially, rather than at its core.
ChatGPT 5.5 generated this card. While it functions correctly and incorporates the required details, this is precisely the issue. The result is indistinguishable from every other dark-mode card produced in the last year. There are no unexpected design decisions, and no element is prioritized; each component receives equal emphasis because the model systematically followed a checklist of specifications.
The solution is not to provide more detailed prompts. Instead, it involves limiting the model’s reference set rather than allowing it to draw from the entire internet. Constraints and references guide the model toward intentional and distinctive outcomes, transforming generic output into a deliberate visual identity. So it helps to look at how constraints function within the creative process.
Artistry in Sampling# #
Originality in music has never meant creating from nothing. Jazz musicians absorb standards, then deviate. Producers can sample, interpolate, tag, warp, and then chop. A remix can stand as an original work that makes its influences explicit. The best ones either stop sounding like their source material or let the source disappear into a larger set of the artist’s influences and proclivities.
AI-assisted design works the same way. You’re not escaping the corpus by ignoring it. You also aren’t the one who shoved it all in there. You’re escaping it by selecting from it intentionally and colliding references in combinations the model hasn’t seen assembled that way. Constraints make the selection deliberate.
The concern about plagiarism is real, especially when a model reproduces a specific work or when a product substitutes for the original. That is different from working inside a recognizable lineage. Copying is the use or recreation of a specific work. Inspiration is absorbing a genre, a tone, a structural pattern. Bebop built an entirely new tradition on the harmonic vocabulary it inherited from swing. That is how subgenres form and evolve. You’re working within a lineage, and reproducing a specific work is a separate act.
These apps live on the local network. No public hosting, user auth to engineer or subscribe to, or deployment pipeline. Access is solved for my uses: tailnet, Hermes agent, or a direct connection from the same machine. As a result, a whole layer of product-design pressure falls away: onboarding, growth loops, account management, public trust signals, pricing pages. With those demoted, the space they would have taken goes to personal features, the ones that only need to make sense for one user.
I think of it as a software bookshelf. Each tool does one thing, lives close to where I work, and doesn’t need to justify its existence to anyone. Building a yt-dlp frontend means I know exactly what’s in it. There’s no binary to audit, no Chrome extension sold to an ad company, no README that hasn’t been updated since 2021.
The trade-off is maintenance, but a 200-line single-page app costs less to maintain than trusting whatever ranks first in a search or, worse, something that uses personalized recommendations to drive you into an information bubble. You learn Photoshop from YouTube tutorials; with AI, the same should extend to more technical tools like yt-dlp, ffmpeg, or agent-driven orchestration. Software as an Appliance, open by default.
Constraints as Creative Decision# #
More rigorous constraints make AI output more distinctive but require creative direction. “Make it look like LimeWire” is more useful than “make it look good.” Naming LimeWire carries implicit design assumptions: a download list, status indicators, chunky controls. Those baked-in semantics are the point, because they give the prompt a concrete payoff beyond generic polish.
Teenage Engineering works this way in hardware. Their Pocket Operators share the same minimal form factor, but each device is constrained to a specific sonic purpose: one for rhythm, one for bass, one for melody. The EP-133 K.O.II extends that idea into a sampler built around destruction and resampling as its core workflow. Same company, same design language, different creative frame locked in by the hardware itself. You’re not choosing what to make; you’re choosing which constraints to work within. Naming a reference in a design prompt works the same way.
ytdlp#
LimeWire had a specific visual grammar: forest green as the primary accent, a chunky transfer list, early-2000s iconography, a toolbar sized for a 1024×768 monitor. Specifying that as a reference doesn’t produce a LimeWire clone; it produces something that shares the grammar without copying any specific product. In that sense, it’s closer to a doodle in my notebook riffing on the aesthetic, fodder to hone my taste.
I can also use open-source tools like yt-dlp instead of building everything myself. This means I choose open projects instead of closed, subscription-based, or locked-down options. Once I saw what yt-dlp could do, I added features for TikTok downloads, MP3 conversions, and background workers.
Clip#
The postal and package service aesthetic works differently. That visual language has a constrained palette (cream, dark type, heavy serif headers) and brutalist information density. Everything important is visible at once, with no wasted elements. Applied to a clipboard tool, it means substance over style. No sidebar collapses. Text remains selectable immediately or curl
’d over tailnet. You get a list of saved items, quick actions, and an archive for soft deletes.
Frame the constraint the same way you would system-prompt an agent instructing it to: “operate within this visual grammar” or use this metaphor as a framework”is analogous tos telling a model it’s a world-class mobile developer. In this dynamic, the model interprets the directive to apply tet grammar to the specific UI problem, thereby contributing structure and consistency. However, true design agency remains with me as the designer; while the AI channels constraints into a tangible direction, I must set and uphold the broader creative standards, carefully weighing the complexity of the visual identity against practical limits on iteration. This process exemplifies a balance between the model’s generative capacity and my own intentional direction.d.
Style vs. Structure# #
There are two things you can borrow from an inspiration: the look and the logic.
The look is surface. The palette, typeface, and border radius. The logic is more diffuse: how information is prioritized, what’s visible by default, how actions are grouped, what the interface assumes the user already knows.
If you only copy the look, you get a theme. If you copy the logic, you get a design language. The best prompts include both, and they don’t have to come from the same source. This matters because it leads to a UI that feels purposeful, not just decorated. Saying “LimeWire-styled” copies the look. Saying “LimeWire-styled, with the same focus on transfer state as LimeWire had for peer count” copies the logic. In a download manager, the key is showing what’s active, queued, and finished all at once. User experience is described through reference.
When the model uses both look and logic, it handles tricky cases consistently. New screens fit in naturally. The AI follows a clear approach instead of just picking from random “good UI” patterns or avoiding a growing list of bad ones.
## In Practice[#](#in-practice)
### sigil[#](#sigil)
The design constraints for sigil were: dark
, keyboard-first
, one level of navigation depth
, TUI-inspired but rendered natively
. No onboarding, hero page, or pretense. Connection state is the primary view, so the payoff is a faster path to what the app exists to do. This helps separate ‘simple’ from sparse or obfuscated.
Those constraints follow from the use case. A niche desktop control app shouldn’t be selling you or babying you. And shallow depth follows the same logic: you open sigil to connect, host, or register. The user flows came from the ergonomics: Unix interoperability, persistent configuration, and surfacing options in the likely setup order.
The AI built within those specifications, backed by technical proof-of-concept. The specific icon set, the sidebar grouping, the status indicator style: all of it falls out of asking, “What would a TUI-native solution use to represent this element?“
maneframe#
The constraint for this blog was 98.css, a CSS library that replicates the Windows 95 UI system. It handles borders, buttons, and scrollbars. It’s a shared design language that readers already recognize.
The structural inspiration came from mapping each section to an existing Windows 95 application metaphor:
- Projects → File Explorer (folders, icon grid)
- Blog → Outlook (inbox list, subject and date columns, newest first)
- About → System Properties (dialog, tabs, system info layout)
Those metaphors came first. The AI built components that fit each one. Because the metaphors are familiar, the result is coherent without needing to explain itself. Projects are numerous and read like a directory. Blog posts are communication, timed. About is static, the same as every System Properties dialog.
The part that isn’t 98.css: the avatar, the typography choices, the content. That’s where the voice lives. The constraint gave the model a design system to work within. The gaps in that system are where the decisions became personal.
Finding the Voice in the Gap# #
The corpus is large and diffuse, with competing ideals and trade-offs baked into the aggregate. An unconstrained AI samples it all and returns the average answer, not the best fit. Constraints narrow down the options. Inspiration helps the model choose certain ideas over others. Together, they focus the design from ‘everything online’ to ‘what fits this style for this problem.’
My personal voice functions as the guiding force that defines the project’s direction, rather than adhering to any prescriptive rules. It emerges from the collision of taste, unique associations, and ongoing technical experimentation: for example, adapting Win95 to a personal blog, employing LimeWire’s design for a down, or integrating postal service aesthetics into a clipboard tool. While these individual references are not inherently novel, it is the deliberate combination, interpretation, and final expression, shaped by personal perspective, that establishes a distinct and cohesive visual identity.
The AI takes care of the technical work. You bring the direction, the taste, and the creative spark.
Sources