I am not a designer. For most of my career, anything that required visual work meant either hiring someone or asking a favor. Tools like Canva helped, and I've used Canva well for what it does. But when AI image generation arrived, I started using it for my blog post headers, and the process of getting a usable image from a raw large language model taught me something I didn't expect: the bottleneck moved from "I can't use design tools" to "I can't write a good enough prompt."

Adobe shipped two things this week that speak directly to that problem. The Firefly AI Assistant hit public beta, and the Adobe for creativity connector launched inside Anthropic's Claude. Both matter. But not for the reasons Adobe's marketing will emphasize.

The Narrow Prompt Is the Whole Point

I've written about this before, and the gap has only gotten wider. When you use a general-purpose large language model to generate an image, you become the art director, the designer, and the prompt engineer simultaneously. You have to specify the style, the composition, the color palette, the aspect ratio, the lighting, what you don't want in the image, and sometimes the negative prompts are longer than the positive ones. You write a paragraph to get a picture. Then you iterate. Then you write another paragraph because the first output had the wrong mood, or the text was garbled, or the composition was off.

That's the LLM image generation experience for a non-designer. You're compensating for the model's lack of creative domain knowledge by stuffing it all into the prompt.

Adobe's approach is the opposite. You write a short prompt. "Blog header, dark background, two contrasting network diagrams, one broken, one connected." The system knows what that means because it has 30 years of creative tool logic behind it. It knows what "portrait retouch" implies across lighting, crop, blur, and tone. It knows what "social asset" means for Instagram versus TikTok versus LinkedIn. You describe the outcome. Adobe fills in the professional knowledge you don't have.

That is the narrow prompt. Short input, professional output, because the domain expertise lives in the tool, not in your head.

Two Launches, One Strategy

The Firefly AI Assistant, now in public beta, is Adobe's creative agent living inside the Firefly app. You describe what you want in plain language. The assistant orchestrates multi-step workflows across Photoshop, Premiere, Lightroom, Illustrator, Express, and more. It draws from over 60 pro-grade tools: Auto Tone, Generative Fill, Remove Background, Vectorize, presets, and a growing list. It includes pre-built Creative Skills for common tasks like batch editing photos, building mood boards, retouching portraits, and creating social variations.

The Adobe for creativity connector does something different. It puts 50-plus of those same pro-grade tools inside Claude. You don't leave the chat. You describe what you want, and the connector orchestrates Photoshop, Illustrator, Firefly, Express, Premiere, Lightroom, InDesign, and Adobe Stock from within the conversation. Sign in with an Adobe account and you get higher usage limits, more tools, and work that saves across sessions.

The strategy is clear. Adobe wants its creative tools to be available wherever you already work, not just inside Adobe's own apps. Firefly AI Assistant is the full experience. The Claude connector is the distribution play.

What You Actually Get That an LLM Can't Give You

I want to be specific about this because the difference is practical, not theoretical. I use both. Here's where they diverge.

When I ask an LLM to generate a blog header image, I get one output. If I don't like it, I re-prompt. If I want to edit part of it, I either describe the edit in another long prompt and hope the model understands, or I download it and open a separate tool. Every session starts from zero. The model has no memory of my style preferences, no understanding of my brand, no concept of what worked last time. I am teaching it everything, every time.

Adobe's system works differently in ways that matter for someone who creates content regularly.

Multi-step workflows from a single prompt. You say "turn this product shot into social assets for Instagram, TikTok, and LinkedIn." The assistant handles cropping, resizing, aspect ratios, and composition for each platform. With an LLM, that's three separate prompts minimum, and you're specifying dimensions manually each time.

Editing after generation. Generative Fill, Remove Background, AI Markup, Precision Flow. These are post-generation tools that let you refine without starting over. LLM image generation is mostly a one-shot process. You get an output and you either accept it or re-roll.

Commercially safe output. Firefly is trained on licensed Adobe Stock content and public domain material. For anyone publishing content professionally, the intellectual property question matters. Most LLM image generators have murky training data provenance. Adobe offers indemnification. That's not a feature for hobbyists. It's a feature for anyone whose content represents a business.

Consistency across sessions. Firefly AI Assistant learns your creative preferences over time. Creative Skills give you repeatable workflows. Your assets save to Creative Cloud storage and are accessible across apps. An LLM conversation is ephemeral. Close the tab and the context is gone.

Third-party model access inside Firefly. This one surprised me. Adobe announced that OpenAI's GPT Image 2 is now available in Firefly, alongside Google's Nano Banana 2, Runway's Gen-4.5, ElevenLabs' Multilingual v2, Kling 3.0, and others. Adobe is becoming a creative model marketplace, not just a single-model tool. You get the generation quality of frontier models with Adobe's editing and workflow layer on top.

This Is Not for Designers. That's the Point.

Professional designers will use Firefly AI Assistant to speed up production work. Fine. But the more interesting market is people like me: content creators, marketers, executives, small business owners who need professional-quality visual assets and have never opened Photoshop voluntarily.

Adobe's own framing confirms this. The blog post lists the target audience as students, content creators, and small business owners. The Claude connector is aimed at people who are already in a chat interface doing other work and want to create something visual without switching contexts. These are not power users. These are people who need a result and don't want to learn a tool to get it.

The narrow prompt is what makes that possible. A designer can write a detailed prompt because they know the vocabulary. They know what "high key lighting with a shallow depth of field and a complementary color palette" means. I don't. I know "clean blog header, dark background, professional look." Adobe's system translates that into the professional vocabulary for me. An LLM makes me learn the vocabulary first.

The Business Question Behind the Beta

Adobe is making a platform bet. Firefly AI Assistant is the owned experience with the full tool set, the learning preferences, the Creative Cloud integration. The Claude connector is the distribution channel that puts Adobe's tools where non-Adobe users already spend time. Both feed the same subscription engine: Creative Cloud Pro or paid Firefly plans for the full experience, with complimentary daily credits during the beta to get people in the door.

The competitive question is whether the narrow prompt advantage holds. LLM image generation is improving fast. OpenAI, Google, and others are getting better at interpreting short prompts and producing usable output. If general-purpose models close the gap on creative domain knowledge, Adobe's advantage shrinks to editing tools and commercial safety. Those matter, but they're harder to sell to a casual user than "just describe what you want."

I don't think that gap closes soon. Creative domain knowledge is deep and specific. Knowing that a LinkedIn carousel needs different composition than an Instagram Reel, knowing that portrait retouching involves seven distinct adjustments in a specific order, knowing that a mood board requires a particular relationship between reference images, that knowledge took Adobe decades to accumulate. An LLM can approximate it. Adobe can execute it.

For now, the narrow prompt wins. And for non-designers who create content professionally, that's the only metric that matters.