In April 2026, Adobe announced the Firefly AI Assistant. According to Adobe’s press release, users can describe the result they want in their own words, and the assistant plans and carries out multi-step work across creative tools such as Photoshop, Illustrator, and Lightroom. Adobe’s documentation also says users can accept or reject the changes it suggests and step in at any point to adjust the direction.

After trying it at the end of May, The Verge’s headline called it “a mediocre design intern.” The review also notes that it’s quite candid about its limits: asked to split objects in a JPG into separate layers, it said it couldn’t, then offered two alternative approaches and explained the pros and cons of each.

The intern comparison is useful. An intern can quickly produce a few directions, but given only “make it more premium,” they can only guess. Where the image goes, what must not appear in it, and what counts as usable: that art direction work is something AI won’t fill in for you. Before generating, write it down as conditions you can check.

Looks good isn’t the same as usable

You type “make it look more premium,” and a few seconds later the image appears, with beautiful lighting and a complete composition. But when you drop it into a website cover, there’s no room for the headline; shrunk to a social preview, the subject is unreadable; and a brand colleague points out a shape that looks like a fake logo.

The problem here isn’t the tool; it’s treating “looks good” as “usable.” Design AI is very good at filling the frame: adding text, icons, interfaces, and decorations to make the image look more complete. What most often blocks production assets is something in the image that shouldn’t be there.

Write four things before generating

Purpose: Is the image for an article cover, a deck’s title slide, an ad, or a product page? Are readers scrolling quickly on a phone or looking closely on a desktop? Should it help them understand a feature, a risk, or a feeling?

Layout constraints: where the headline goes, how much space to leave on the sides or top, and whether the key point is still visible as a mobile thumbnail.

What must not appear: no text, no logos, no fake interfaces, no recognizable people, nothing imitating a specific brand closely enough to confuse.

What counts as usable: the main message is clear in the actual layout, the key point is recognizable when shrunk, there are no fake letters or stray symbols, and the style doesn’t clash with the existing brand.

With these four things written down, AI output has a standard to check against. Without them, you can only keep sending “make it look better” and generate more images that create more hesitation. The same goes for revision requests: instead of “more polished,” say which version to compare against and what to change; see AI Revisions Go in Circles Because “More Polished” Can’t Be Compared.

Advertisement

Hand exploration to AI, but have a person revise before launch

Design AI is well suited to early exploration: finding cover directions, trying a few moods, making drafts for internal discussion, generating compositions for low-risk articles or decks. Speed matters most here, because the goal is to see options quickly.

But if an asset will appear in a production ad, a brand key visual, a paid campaign page, a product interface, or anywhere that requires precise text and confirmed rights, generated results can’t go live directly. A person still needs to revise it, integrate the layout, check brand guidelines, and confirm usage rights.

Clear conditions don’t guarantee every image works

Writing conditions clearly reduces guesswork, but it doesn’t guarantee every image will work. Some judgments, such as whether an image fits the brand’s tone or might mislead readers, still need someone who knows the brand to look.

If the design will next be handed to engineers to build, what to check at design handoff is a separate question; see Once a Figma Canvas Can Run Code, Mark What’s Intent and What’s Experiment at Handoff. For leaving records that let colleagues redo the work, see Charts Made With AI Should Be Reproducible by a Colleague Months Later.

AI handoff card

I'm about to ask a design AI for an image, and I want to write the art direction as checkable conditions first instead of just saying "make it look good."
The method comes from this article: https://boosterminiclass.com/en/posts/design-ai-assistants-need-human-art-direction/
Its view: design AI can carry out a description, but it doesn't know where the image goes, what must not appear, or what counts as usable. Write these down before generating; the first good-looking image is only a draft.

First ask me: where the image will go, who will see it, what it should help readers understand, and whether there are existing brand guidelines.
Then:
1. Turn my answers into a generation brief covering purpose, layout constraints (headline position, white space, mobile thumbnail), elements that must not appear, and the conditions for usability;
2. Based on the brief, write instructions for the design AI, with style adjectives at the end rather than the beginning;
3. List what to check after the first round, such as whether the main message is clear in the layout, whether the key point is recognizable when shrunk, and whether there are fake letters or logo-like shapes;
4. State whether this image is an "exploratory draft" or "headed for launch," and if the latter, list the human revisions and confirmations still needed.
Don't decide the style direction for me; ask about the conditions first.
Advertisement

Share

Share this mini class

If this lesson helps untangle a work bottleneck, share it with someone deciding how to use AI.

References