Prompt Engineering Was Never the Answer for Campaign Visuals

Sixty-seven percent of marketers now create or enhance images with AI, up from 39% in 2024, according to the Social Media Examiner 2026 AI Marketing Report (n=681). Yet in the same survey, 61% agree that AI content feels inauthentic or off-brand. Adoption went up. Trust in the output did not keep pace.

That gap has a cause that rarely gets named: we spent three years treating image generation as a writing problem. Learn the magic words, stack the modifiers, save the prompt library, and the results will come. For one hero image, that works well enough. For a campaign, it was never going to.

If you run a creative agency, a design studio, or the visual side of an e-commerce or fashion brand, this post is for you. It argues that prompt engineering was a workaround for a missing feature, shows what the major tools quietly shipped to replace it, and explains what to do on Monday morning. We built Rainfrog inside a design agency, so the view here is from production, not from a demo reel.

Table of Contents

• What Was Prompt Engineering Actually For?
• Why Prompts Can't Carry a Campaign
• What the Big Tools Shipped Instead of Better Prompts
• The Hidden Cost of Prompt Libraries
• What Should Replace Prompting in a Campaign Workflow?
• How to Audit Your Team's Prompt Dependence
• Where Prompts Still Earn Their Place
• Frequently Asked Questions
• Key Takeaways

What Was Prompt Engineering Actually For?

Prompt engineering was a workaround for tools that only accepted text. When the only control surface is a sentence, you learn to write better sentences. The skill was real, but it compensated for a missing interface rather than describing how creative work should be done.

The wider labour market has already drawn the same conclusion. In May 2025, Fast Company reported that prompt engineering had mostly stopped being a standalone job. Malcolm Frank, CEO of TalentGenius, put it this way: "It's turned from a job into a task very, very quickly." Aline Lerner of Interviewing.io added that online discussion of the role "was probably much bigger than the head count."

For image work, the lesson is sharper. A task that an art director once did with a mood board, a product sample and a conversation was squeezed into a text box. Of course it took tricks. Our earlier post on why prompt engineering is the wrong approach for campaign imagery covers the mechanics. This one is about the pattern behind it.

Why Prompts Can't Carry a Campaign

A prompt describes an image. A campaign is a set of images that must agree with each other. Text is a poor way to enforce agreement, because the model re-reads your words from scratch on every generation, and small differences in interpretation compound across a set.

The same words, different shoot. Write "editorial lighting, muted palette, linen jacket" twenty times and you get twenty plausible interpretations. Each is fine alone. Together they look like twenty different photographers. We unpack this failure mode in what campaign visual consistency is and why AI usually gets it wrong.

Specific products don't survive translation. You can describe "a tan trench coat with a storm flap." You cannot describe your own trench coat, with its exact collar, stitching and button spacing, in words. The model returns a trench coat, not your trench coat.

Brand governance sits outside the prompt. Jasper's 2026 State of AI in Marketing report (n=1,400, published January 2026) found that 91% of marketing teams use AI, but that governance is now the top challenge, with blockers from legal, compliance and brand review rising 3.4x year over year. A paragraph of prose is a weak container for brand rules that reviewers need to check.

Knowledge lives in one person's head. The prompt that works is usually the property of whoever wrote it. When that person is on holiday, the look changes. That is a staffing risk, not a creative one.

What the Big Tools Shipped Instead of Better Prompts

Every major image tool has added a way to show the model what you mean instead of telling it. That is the clearest evidence that text alone was not enough. Reference images, style anchors and trained brand models all move control from words to assets.

Midjourney: reference images. Midjourney's Omni Reference lets you carry a character, object or creature from one image into a new creation. The documentation is candid about the limits: it accepts one reference image per prompt, still needs a text prompt alongside it, and notes that fine details such as freckles or clothing logos may not match the reference exactly. Its separate Style Reference captures color, medium, texture and lighting from an image, and likewise requires text to be used with it.

Adobe: models trained on your assets. Adobe's Firefly Custom Models are described as "fine-tuned Firefly generative AI models that you train with your own assets," aimed at on-brand images across web, paid media, email and social. Adobe's page makes consistency claims but publishes no measured data on how consistent the output actually is, so treat it as a vendor position.

Google: character persistence. When Google introduced its image model in the Gemini 2.5 Flash Image announcement (August 2025), the headline capabilities were keeping a character or object's appearance across prompts and edits, and following visual templates. Google also said it was still working on "even more reliable character consistency."

Read those three together. A text-to-image company, a design-software company and a search company all arrived at the same fix: stop describing, start supplying. None of them presented it as "a better way to prompt." They presented it as a way to need prompts less.

The Hidden Cost of Prompt Libraries

Prompt libraries feel like a systems answer, but they mostly formalise the workaround. A shared document of 40 prompts needs maintenance every time the model updates, because the same wording can produce a different look after a version change. Midjourney's own documentation flags this for style codes, noting they may not give the same look after a major version update.

There is also a time cost that does not show up on invoices. Agencies bill for the finished asset, but the hours go into iterating on wording: adjust the lighting phrase, re-roll, adjust the lens phrase, re-roll. That loop is invisible in the final file and expensive in the schedule.

The teams that notice this first are usually the ones producing volume: e-commerce catalogues, seasonal drops, social calendars. For them, a library of prompts is a fragile stand-in for what they actually need, which is a repeatable system. We describe what that looks like in how to set up an AI visual production workflow for a design agency.

What Should Replace Prompting in a Campaign Workflow?

The replacement is a set of reusable inputs: the product, the character, the style and the environment, saved once and combined for every shot. The campaign's look then lives in assets your team can inspect, approve and reuse, not in a sentence someone has to remember.

This is the model Rainfrog is built around. You pick the pieces of the shoot, and the platform keeps them consistent across the set. The walkthrough is in how Rainfrog generates campaign visuals without a single prompt, and the broader definition is in what campaign-level AI image generation actually means.

Whatever tool you use, judge it on four questions:

1. Can I save a product, character, style or environment once and reuse it?
2. Does the second image match the first without rewording anything?
3. Can a colleague produce the same look without asking me for my prompt?
4. Can a brand or legal reviewer approve the inputs, not just the outputs?

If the answer to most of these is no, you are still prompting your way through the campaign, whatever the interface looks like.

How to Audit Your Team's Prompt Dependence

A one-hour audit will tell you how much of your look lives in people's heads. You don't need special software, just last month's work and a few honest questions.

Step 1: Pull one finished campaign. Choose a set of 8 to 12 visuals that shipped to a client.

Step 2: Find the prompt behind each image. If you can't, that is the first finding. Note how many were written by the same person.

Step 3: Count the rerolls. For each final image, estimate how many generations it took. High numbers on product shots usually point to detail the prompt could not hold.

Step 4: Mark what a reference could have fixed. Circle every correction that was about a specific product, face or style. Those are input problems, not wording problems.

Step 5: Decide what becomes a saved asset. Anything repeated across campaigns, such as a signature background or a recurring model, should move out of prose and into a reusable input.

The result is usually a short list of assets that, once saved, remove most of the prompting you were doing by hand.

Where Prompts Still Earn Their Place

Prompts are not useless. They are the right tool for exploration: testing a mood, trying an unusual composition, or getting a first idea on the board before anyone commits to a direction. Both Midjourney's and Google's documentation still pair references with text, and that is sensible.

The distinction we'd draw is between ideation and production. Prompting is a fine way to find a look. It is a fragile way to reproduce it forty times. Use text to explore and inputs to deliver.

It is also worth being clear about what we can't promise. Reference-based approaches reduce drift, but none of the vendors above claim to eliminate it, and Google and Midjourney both document remaining gaps. Expect to review outputs, and plan your approval step accordingly.

Frequently Asked Questions

Is prompt engineering dead for image generation?

Not dead, but demoted. It remains useful for exploring ideas, yet the major tools now rely on reference images and trained models for repeatable results. As Fast Company reported, the standalone job has largely become an everyday task.

Why do my AI images look different even with the same prompt?

The model reinterprets your wording on every generation, so small variations accumulate across a set. Model updates can also change how the same words render. Anchoring the output to saved inputs, as covered in our guide to consistent brand visuals, reduces that drift.

Do reference images fully solve consistency?

No. Midjourney's documentation says fine details like logos may not match a reference exactly, and Google says it is still improving character consistency. References narrow the variation considerably, but production teams should still review outputs.

What should an agency standardise instead of prompts?

Standardise reusable inputs: products, characters, styles and environments, plus the approval steps around them. Our post on adding AI visuals to your agency workflow walks through the rollout.

Is a no-prompt tool right for solo creators too?

Often yes, because the time saved on rewording and rerolling matters most when one person does everything. See the pricing page for plans.

Key Takeaways

• Prompt engineering was a workaround for text-only interfaces, not a lasting creative skill.
• Prompts describe single images; campaigns need images that agree with each other.
• Midjourney, Adobe and Google have all moved toward references and trained models, which signals text alone was not enough.
• Prompt libraries formalise the workaround and break when models update.
• Use prompts for exploration and saved inputs for production, and keep a human review step either way.

Want to see what a campaign built from inputs looks like? Try Rainfrog and see the workflows.