The Death of the Asset Library

Reading Time: 11 minutes

---

“AI can build almost anything. What it can’t do is determine what’s worth building.” — Figma, 2026 AI Report

---

The Death of the Asset Library

Why Prompting is Replacing the Stock Website

---

The Great Asset Hunt

When every project began with a search bar

For years, a fairly ordinary design job could begin with an hour of searching. The direction was approved, the brief was clear enough, and somewhere inside Shutterstock, Envato Elements, Adobe Stock or the Figma Community there was probably an image, icon, mockup or component that could be made to work.

Finding it was another matter.

The illustration might have the right composition and completely wrong colours. The photograph worked except for the background. A mockup could be almost perfect until you put the client’s product into it and realised the angle made the thing look strangely small. You searched, downloaded, edited, cropped, recoloured and occasionally rebuilt half the asset anyway.

Stock libraries were useful because very few projects can justify creating every photograph, icon or mockup from scratch. They solved a real production problem. They also trained designers to work around whatever happened to exist.

You can still see the result on the web. The same office photographs turn up in different industries. Familiar illustration styles appear under unrelated brands. There is always another photograph of several suspiciously cheerful people pointing at a laptop.

For a long time, none of this was particularly unusual. You searched first, found something close enough and made the design accommodate it.

Generative tools change that order. The question is slowly moving away from “Where can I find something like this?” and towards “What do I actually want this to look like?”

---

From Searching to Summoning

Why find the asset when you can describe it?

Generation is no longer something designers have to leave their main tools to experiment with. Adobe has been putting Firefly-powered and partner-model features directly into Creative Cloud. Photoshop now includes Generative Fill, Generative Expand and Generate Image. Illustrator has tools including Text to Vector Graphic, Concept to Vector, Generative Shape Fill and Generative Recolor.

That placement matters more than it first appears. A designer does not necessarily have to stop working, open another service, generate something, export it and drag it back into the job. Generation starts behaving less like a separate destination and more like another thing Photoshop or Illustrator happens to do.

Midjourney has been moving from the other direction. What many people first used as a fairly simple prompt-and-generate tool now has a browser editor, reusable aesthetic profiles and Moodboards. Its Personalization system can carry aesthetic preferences into later generations, so each prompt does not have to begin from zero.

A search like “isometric technology illustration blue” used to be enough to start digging through results. Now the same requirement can be treated more like an art-direction brief:

“Create an isometric vector illustration of a compact AI data centre viewed from slightly above. Rounded geometric forms, no outlines, four-colour palette using deep navy, warm cream, electric cyan and muted coral. Minimal shadows, no gradients, clean editorial-tech aesthetic, generous negative space on the right for headline copy and consistent visual weight.”

The first result may still be wrong. It may need references, refinements, several variations and the occasional increasingly specific instruction because the model has decided your “minimal shadow” means something entirely different.

The important part is where you start. You are no longer choosing the closest object somebody else already made. You are describing the object you wanted in the first place.

Once that becomes normal, the asset library starts to feel different. You may still need one, but it is no longer the only place where the raw material has to come from.

---

The Library Becomes a Prompt

From choosing what exists to defining what should exist

Traditional asset libraries work by giving you enormous amounts of material and hoping the useful thing is somewhere inside. Another ten thousand photographs means another ten thousand possibilities to search through.

Generative systems do not need every variation to exist beforehand. If the variation is not there, the system can attempt to make it when you ask.

The idea is spreading outside imagery. Figma Make puts prompt-driven creation alongside the components, prototypes and design systems designers already use. That becomes more interesting when generation can work with an existing system instead of inventing another glossy card layout with no relationship to the product around it.

v0 by Vercel has been doing something similar on the development side. The product has moved beyond generating isolated UI examples towards a broader application-building environment. Vercel reported in February 2026 that more than four million people had used v0 since general availability. Its work around AI-powered prototyping with design systems also points towards generated interfaces that can inherit existing components instead of repeatedly starting with generic AI styling.

Old Search Workflow	New Prompt Workflow
Search for the closest suitable asset	Describe the asset you actually need
Download a generic template	Generate from existing brand context
Recolour, crop and reconstruct	Refine the generated direction
Repeat the process for another format	Create controlled variations
Asset library stores the possibilities	Brand rules shape the possibilities

The difference becomes much easier to see once a project has more than one output. A campaign that needs twelve formats does not necessarily need twelve separate searches. One approved direction can be adapted into different crops, compositions and supporting visuals. A rough interface idea can become something editable before anyone has spent an afternoon digging through UI kits looking for a screen that happens to be close.

Stock companies have noticed the same shift, which makes the idea of their sudden disappearance rather unlikely.

---

Not Quite Dead Yet

Reports of the stock website’s death have been slightly exaggerated

“The death of the asset library” makes a better headline than “the asset library changes quite a lot,” but Shutterstock is not disappearing tomorrow. Neither are Adobe Stock, Envato or the specialist libraries designers use every day.

Existing assets still have one very useful advantage: you know what you are getting.

A licensed photograph is already there. A professionally built mockup can be opened and inspected. A proper icon family generally contains icons that look as though they were designed by the same person, which is not something every generative system manages consistently.

Shutterstock itself makes the change easier to see. In 2026, the company expanded AI editing within its search experience, allowing users to work with generated or edited material alongside its traditional library. Shutterstock describes the approach as combining human-created imagery with AI-assisted editing rather than treating the two as separate camps. Its AI image editing experience points in that direction.

That kind of hybrid workflow feels much closer to what designers are likely to encounter in the near term. You might begin with a licensed photograph, extend the background with Firefly, alter the composition, create several campaign crops and finish with something quite different from the asset you originally searched for.

The photograph did not disappear. What changed was the amount of work it could participate in afterwards.

There is another consequence. Once libraries and generators can both produce more material than anybody has time to inspect, finding enough options stops being the problem.

---

The Infinite Choice Problem

More options do not automatically manufacture better taste

Search results used to end eventually, or at least your patience did. A designer might inspect twenty useful options, reject most of them and move on.

Generative systems make stopping less obvious.

Midjourney can try another direction. Firefly can regenerate one part of an image. Figma can produce another interface idea. v0 can rebuild the idea as working UI. There is nearly always another version available if you are willing to ask for it.

Art Director Urmila Nandan described her experience with AI in creative-direction work simply: “AI doesn’t reduce the effort  it just shifts it.”

That feels accurate. Generating another option may take less effort, but somebody still has to look at it. Then compare it with the previous one. Then decide whether the interesting detail in version thirty-seven is worth another round.

A creative director who once reviewed four directions can now receive forty without anybody spending ten times as long producing them. The production bottleneck shrinks. The judgement bottleneck does not.

Figma’s 2026 AI Report gives the point some useful context. Based on 8,403 survey responses and 639 qualitative interviews with designers, developers and product managers, Figma found that 90% of respondents considered design at least as important as it was before AI, while nearly six in ten said it had become more important. Figma obviously benefits from a healthy design industry, but the result still raises an interesting question. If making another thing becomes cheap, deciding whether that thing deserves to exist becomes a larger part of the job.

This becomes much more serious once the work has to remain recognisably part of the same brand. Producing a hundred assets is easy enough to imagine. Producing a hundred assets without slowly turning one identity into a collection of distant relatives is harder.

---

The Brand DNA File

Because a logo and three hex codes were never a personality

Most traditional brand guides were written for people. They showed the correct logo, colour values, typefaces, spacing rules and examples of how everything should appear. Somewhere near the back there was often a paragraph describing the brand as “bold but approachable” or “playful yet professional,” which a human designer could usually translate into something useful after seeing the actual work.

Software needs more detail.

If a generative system is expected to produce layouts, campaign imagery, interface components, presentation graphics and social assets without wandering away from the brand, it needs more than a mood and three approved adjectives.

That is where the idea of a Brand DNA File becomes useful. It does not need to be one literal file, and very few companies need to train a foundation model of their own. It is better understood as the information a generative system would need in order to make sensible creative decisions: design tokens, typography rules, colour values, components, preferred compositions, photographic treatments, illustration references, terminology, accessibility rules, prompt examples and examples of what the brand absolutely should not become.

We can already see parts of this taking shape. Figma is bringing generation closer to existing product and design-system context. Vercel has argued that generated prototypes need access to existing components and their behaviour if the result is going to feel like the real product. Midjourney’s Personalization and Moodboards tackle a similar problem from the visual side by allowing aesthetic context to persist instead of rebuilding it in every prompt.

An agency handover could therefore become much broader than a brand-guidelines PDF. The familiar guide may still be there, but beside it could sit the Figma library, design tokens, approved components, reference imagery, prompt examples, writing rules and reusable instructions for whatever systems the client uses later.

Large organisations may eventually build tightly controlled generative environments around their approved material. Smaller companies can get surprisingly far simply by giving general-purpose systems better context.

The instructions could become very specific. Illustrations use rounded geometry. No photorealism. Shadows stay subtle. Headlines are short. The primary colour works on backgrounds but not body copy. People should look candid rather than posed. Interface cards use a particular radius. “Revolutionary” is not allowed unless somebody has genuinely revolutionised something.

At that point the prompt only needs to describe what is changing.

A marketer might ask for “a launch visual for our new analytics feature”. The system would already have the visual references, components, exclusions and language rules needed to understand what kind of company is asking.

That changes the thing an agency is being paid to build. Producing the first set of assets still matters, but studios may also be asked to create the system that keeps later work recognisable.

The deliverable starts to include the rules behind the work: components, references, constraints, examples and enough context for future generation to stay within the same visual world.

The old brand guide explained the identity. The next version may also need to make that identity usable by software.

In practice, that begins to look less like a folder full of finished assets and more like a creative operating system.

Which leads to the question designers have been asking since generative tools became good enough to be uncomfortable: if clients can make the icon, mockup or landing page themselves, what exactly are they still hiring a designer to do?

---

What Happens to the Designer?

Execution gets cheaper; decisions get more expensive

Whenever software removes part of a creative process, somebody eventually asks what is left for the person who used to do it.

Design is going through that argument now. If an AI can draw an icon, produce a layout, alter a photograph or create seventeen campaign sizes without complaining about the TikTok version somebody remembered at the end, some production work is obviously going to change.

A tool that can generate ten credible directions quickly alters the economics of work that previously took hours to explore. v0 can turn descriptions into working interface concepts. Figma Make can speed up prototyping. Firefly can create and modify imagery inside Adobe’s existing environment.

Pretending none of that saves time would be difficult.

Production, however, was only part of what a good designer was doing.

Somebody still has to understand what the company is trying to communicate. Somebody has to notice that an attractive direction belongs to another brand entirely, or that a perfectly executed prompt has produced exactly what was requested and the request itself was a bad idea.

That kind of work looks more like art direction, editing, curation and system design. If agencies become responsible for the rules that generate later work, the ability to make another variation matters less than the ability to recognise which variation should survive.

A designer may receive ninety-nine plausible options and keep one.

The software can help with the ninety-nine. The uncomfortable part is that somebody still has to know which one deserves to be number one.

Generation makes competent execution easier to obtain. A recognisable point of view remains much harder to manufacture.

---

Reading List: For the Prompt-Curious

Useful reading for the point where “make it more premium” stops being a sufficient creative brief.

• John Maeda – How to Speak Machine: A useful bridge between familiar design thinking and the computational systems increasingly sitting beside it.

---

• Don Norman – The Design of Everyday Things: The tools keep changing. The awkward ways people understand, misunderstand and use designed things are considerably more persistent.

---

• Ethan Mollick – Co-Intelligence: A practical look at working with generative AI without treating it as magic or assuming the robots have already won.

---

• Paul Rand – Design, Form, and Chaos: Worth returning to when new tools make it tempting to confuse faster production with better design.

---

• Kate Crawford – Atlas of AI: A look underneath AI at the labour, infrastructure and material systems that disappear very easily behind a clean interface.

---

Final Thought

The library becomes infinite; taste stubbornly refuses to scale with it

Stock websites gave designers enormous shelves of things that already existed. Generative design changes the shelf.

The asset no longer has to be sitting somewhere waiting to be found. It can be produced when the job needs it, changed around the space it has to occupy and regenerated when the brief changes fifteen minutes before approval.

That does not make stock photography obsolete. It does not make templates pointless, and it certainly does not mean every designer now has to communicate with software entirely through prompts.

The more likely workflow is messy and mixed. A project might use licensed stock, original photography, AI generation, manual illustration, traditional editing, design systems and code without anyone feeling particularly obliged to decide which category the final asset belongs to. Shutterstock’s move towards combining searchable library material with generative editing is already making the distinction less tidy.

What changes is where some of the value moves.

If images and variations become extremely easy to produce, simply owning a larger pile of them matters less. The useful part becomes the set of decisions that keeps those outputs coherent: how the brand looks, how it speaks, what it avoids, which references matter, which components are approved and where the boundaries sit.

Agencies may spend less time filling libraries one asset at a time and more time defining the conditions under which new assets can be made without gradually erasing the identity they belong to.

Perhaps the asset library is not disappearing. It is becoming less like a warehouse and more like a set of instructions for making the next thing when somebody needs it.

The shelf is still there. It just no longer has to be stocked in advance.

---

References

---