Skip to content

AI Image Generation and Design · Consistency across a campaign · lesson 11 of 18 · 7 min

Style systems and campaign prompt libraries

From one-off prompts to a system

When a brand uses AI imagery regularly, individual prompting becomes a bottleneck and a consistency risk. A style system — documented style specs, prompt snippets, references and rules — lets anyone on the team produce on-brand images quickly.

Components of an AI style system

AI style system
1. Style specs      Medium, lighting, palette, texture, composition per image type
2. Prompt snippets  Reusable text blocks: [STYLE_CORE], [LIGHTING], [CONSTRAINTS]
3. Reference sets   Approved style refs, character refs, product photos
4. Settings         Tool, model, aspect ratios, reference strengths, upscale settings
5. Templates        Layout templates with image zones and type
6. Do/don't gallery Approved and rejected examples with reasons
7. QA checklist     Artifacts, brand fit, rights, bias, disclosure
8. Change log       Updates when models or brand rules change

Modular prompt snippets

Break prompts into reusable modules:

[STYLE_CORE]   = "flat vector illustration, soft rounded shapes, subtle paper grain,
                  limited palette of navy, mint and warm sand"
[LIGHT]        = "soft even lighting, no harsh shadows"
[COMPOSITION]  = "generous negative space in the top third for a headline"
[CONSTRAINTS]  = "no text, no logos, no watermarks"
[PEOPLE]       = "diverse, everyday people, simple faces without detailed features"

Prompt = [STYLE_CORE] + scene description + [PEOPLE] + [LIGHT] + [COMPOSITION] + [CONSTRAINTS] + aspect ratio

Only the scene description changes per asset. Store snippets where the team can copy them — a shared document, a prompt manager, or a brand kit feature in your design tool.

Image types within one campaign

A campaign often needs several image types, each with its own spec:

| Image type | Spec focus | |---|---| | Hero visual | Highest quality, strong composition, space for headline | | Supporting scenes | Same style, simpler compositions | | Backgrounds/textures | Low detail, brand colors, subtle | | Icons/spot illustrations | Consistent line/shape language, transparent backgrounds | | Thumbnails | High contrast, bold focal point, readable at small size |

The do/don't gallery

Words cannot capture everything. A visual gallery of approved and rejected outputs, each with a short reason ("Rejected: too glossy, off-palette purple"; "Approved: correct grain and palette"), is one of the fastest ways to align a team or a client.

Aligning AI style with brand guidelines

The AI style system should sit inside the brand guidelines, not beside them:

  • Use brand palette names and hex references.
  • Match the brand's photography or illustration style rules.
  • Apply brand accessibility standards (text contrast, inclusive depiction).
  • Follow brand rules on disclosure and where AI imagery is not allowed (for example, customer testimonials).

Managing model changes

Tools update models frequently. After an update:

  1. Re-run a small benchmark set of standard prompts.
  2. Compare with the approved gallery.
  3. Adjust snippets or reference strengths if the style has drifted.
  4. Record the change in the change log.

Pinning a specific model version, where the tool allows, can keep results stable during a campaign.

Worked example: an agency's system for a fintech client

  • Image types: hero illustrations, feature spot illustrations, social backgrounds.
  • Snippets: [STYLE_CORE], [PEOPLE] (inclusive depiction across ages, genders and backgrounds relevant to the client's markets in the UK, UAE and Pakistan), [CONSTRAINTS].
  • References: six approved hero images; icon set drawn by a human designer.
  • Rules: no AI imagery of "customers" in testimonials; financial claims never implied by images; all text added in design tool.
  • Benchmark: eight prompts re-run monthly and after model updates.

Common mistakes

  • Every team member writing prompts from scratch.
  • Style systems that exist only in one person's head.
  • No benchmark after model updates, causing silent drift.
  • An AI style that conflicts with the brand's existing photography or illustration rules.

Rolling a style system out to a team

A style system only helps if people use it. Keep it in one shared location, name snippets clearly, and show a before-and-after example for each so new team members understand the effect. Run a short onboarding session in which each person generates three images with the snippets and compares them with the gallery. Collect questions in a shared document and update the system monthly. For freelancers and partners, share a lighter version that contains the snippets, references and rules they need without exposing confidential client material.

Hands-on: a prompt library your team will use

Store the library where people already work — a shared doc, Notion or Airtable base, your DAM, or the brand-kit features in Canva, Adobe Express or Figma. Each entry needs four things: the snippet, a sample image, when to use it, and who owns it.

LIBRARY ENTRY
id:          fintech-spot-illustration-v2
image type:  Spot illustration (blog, carousel covers)
assembled prompt:
  [STYLE_CORE] + "a young couple reviewing a budget on a tablet at a kitchen
  table" + [PEOPLE] + [LIGHT] + [COMPOSITION] + [CONSTRAINTS] + "4:5"
references:  style_ref_01-06 (medium strength)
approved example: /approved/spot/2026-08-couple-budget.png
don't:       coins raining, piggy banks, dollar-sign clichés
owner:       Art director (A. Khan)   last tested: 2026-09 on [model/version]

A tiny script can assemble prompts from snippets so nobody retypes them (use it with any tool's API or paste the output into the app):

SNIPPETS = {
    "STYLE_CORE": "flat vector illustration, soft rounded shapes, subtle paper grain, "
                  "limited palette of navy, mint and warm sand",
    "PEOPLE": "diverse, everyday people, simple faces without detailed features",
    "LIGHT": "soft even lighting, no harsh shadows",
    "COMPOSITION": "generous negative space in the top third for a headline",
    "CONSTRAINTS": "no text, no logos, no watermarks",
}

def build_prompt(scene: str, ratio: str = "4:5") -> str:
    order = ["STYLE_CORE", "PEOPLE", "LIGHT", "COMPOSITION", "CONSTRAINTS"]
    parts = [SNIPPETS["STYLE_CORE"], scene] + [SNIPPETS[k] for k in order[1:]]
    return ". ".join(parts) + f". Aspect ratio {ratio}."

print(build_prompt("a shop owner in Karachi checking sales on her phone"))

Model updates: a change-log routine

When a vendor ships a new model version, run your library's regression set — the same eight to twelve prompts with the same references — and compare against the approved examples. Record the result:

| Date | Model change | Regression result | Action | |---|---|---|---| | 2026-07 | Tool default model updated | Palette warmer; line weight heavier | Added "cool neutral palette" to STYLE_CORE; re-approved 10 samples | | 2026-09 | New editing model in design suite | No change to generation; better inpainting | None; noted in library |

(Illustrative entries.) Until the regression set passes, keep the team on the previous model where the tool allows it.

Summary

Build an AI style system with specs, modular snippets, references, settings, templates, a do/don't gallery, a QA checklist and a change log; align it with brand guidelines; and benchmark after model updates.

Video lecture: Style systems and campaign prompt libraries

Lecture coming soon · 10 chapters · about 8 minutes. Read the full transcript below.

  1. Style systems and prompt libraries
  2. Why a system
  3. Eight components
  4. Modular snippets
  5. Worked example 1: a solo creator's mini system
  6. Worked example 2: an agency's fintech library
  7. Watch me do it: entry + regression set
  8. Going further
  9. Common mistakes
  10. Recap and try this now

Lecture transcript

Style systems and prompt libraries

Here's what happens in most teams that start using AI imagery. One designer gets really good at it. Their prompts live in their head and their chat history. Then they go on vacation, and the brand's images suddenly look different. Or the team grows, and five people produce five slightly different styles. The fix isn't a better prompt. It's a system. In this lecture, you'll learn the components of an AI style system, how to build modular prompt snippets, how to organize a prompt library your team will actually use, and how to handle model updates without breaking your look. By the end, anyone on your team should be able to produce an on-brand image in minutes.

Why a system

Why build a system? Because consistency at scale is a team sport. Brand guidelines have always done this for logos, colors and fonts. A style system does it for AI imagery. It turns tacit knowledge into shared, documented building blocks. Think about a restaurant kitchen. The head chef doesn't cook every plate. They write recipes, set standards, and check the pass. That's how ten cooks produce the same dish. Your style system is the recipe book. And there's a bonus. It makes onboarding freelancers and new hires much faster, and it gives clients confidence that the look will hold even as the team changes.

Eight components

Here are the eight components. Style specs for each image type. Prompt snippets, reusable text blocks for style, lighting, composition and constraints. Reference sets, including style, character and product references. Settings, meaning tool, model, aspect ratios, reference strengths and upscale settings. Layout templates with image zones and type. A do and don't gallery with approved and rejected examples and reasons. A QA checklist. And a change log for when models or brand rules change. You don't need a fancy tool. A shared document works. But every component must exist somewhere the whole team can find, and each one needs an owner.

Modular snippets

Let's look at modular snippets, which are the engine of the system. Instead of writing whole prompts, you break them into named blocks. Style core: flat vector illustration, soft rounded shapes, subtle paper grain, a limited palette of navy, mint and warm sand. People: diverse, everyday people with simple faces. Light: soft even lighting. Composition: generous negative space in the top third. Constraints: no text, no logos, no watermarks. Then a prompt is just: style core, plus the scene, plus people, light, composition, constraints, and the aspect ratio. Here's the key idea. Only the scene description changes per asset. Everything else is locked, tested and owned.

Worked example 1: a solo creator's mini system

Worked example one, simple. A solo creator who posts education carousels three times a week sets up a mini system in one afternoon. Three snippets: style core, light and constraints. One layout template for covers. A small gallery with four approved and four rejected images, each with a one-line reason, like rejected, three-D rendering breaks our flat style. Now, when she needs a cover, she writes one sentence for the scene and pastes it between the snippets. Her time per cover drops, and more importantly, her feed looks like one creator made it, even when she experiments with new topics. A system doesn't have to be big to work.

Worked example 2: an agency's fintech library

Worked example two, a business scenario with illustrative details. An agency handles a fintech client that needs about eighty images a month across four image types: spot illustrations, hero scenes, app screen backgrounds and social covers. The agency builds a library in a shared database. Each entry has an ID, the assembled prompt, the reference set, an approved example, a don't list, like no coins raining and no piggy banks, an owner and a last tested date. Freelancers get view access and a thirty minute onboarding. Two months later, the tool updates its default model and the palette shifts warmer. Because the agency has a regression set, they spot it on day one, adjust the style core, and re-approve samples before any client asset changes.

Watch me do it: entry + regression set

Watch me do it. I'm setting up a library entry and a regression test. First, I create the entry for spot illustrations. I paste the assembled prompt, attach the six style references, and link one approved example. I add the don't list and my name as owner. Next, I build the regression set: ten prompts that cover our main image types, each with its approved result. I save them in a folder called regression. Now, whenever the tool announces a model update, I run those ten prompts, place the new results next to the approved ones on a board, and check palette, line weight, texture and people. If anything drifts, I adjust the snippets and write the change into the log.

Going further

If you're comfortable with a little code, you can take it further. A tiny script can assemble prompts from snippets, so nobody retypes the style core and nobody introduces typos. The lesson includes a short Python example that does exactly that. You can paste its output into any tool, or feed it into an image API. Some teams go further with node-based AI workflow tools, like Figma Weave, or automation platforms, where a design system component and a snippet feed a generation step automatically. That's powerful, but the same rule applies. The system, meaning snippets, references and review, is what keeps the look consistent. Automation just runs it faster.

Common mistakes

Common mistakes. Keeping prompts in personal chat histories. Building a giant library nobody can navigate. Writing snippets that contradict each other, like minimal in one and highly detailed in another. Having no owner, so the library goes stale. Skipping the do and don't gallery, which is often more useful than the words. And being surprised by model updates. Treat every model update like a software update in your stack. Test before you roll it out. And align the AI style system with the brand guidelines, so the palette, typography and photography rules match what's already approved.

Recap and try this now

Let's recap. A style system turns one person's skill into a team capability. It has eight components: specs, snippets, references, settings, templates, a do and don't gallery, a QA checklist and a change log. Snippets make prompts modular, so only the scene changes. A library entry needs a prompt, references, an approved example, a don't list, an owner and a last tested date. And a regression set protects you from model updates. Try this now. Write four modular snippets for a brand, assemble prompts for three different image types, and start a do and don't gallery with at least two approved and two rejected examples, each with a reason.

Key takeaways

  • A style system — specs, snippets, references, settings, templates, gallery, QA and change log — makes AI imagery a team capability.
  • Modular snippets lock tested style decisions so only the scene description changes per asset.
  • Library entries need an assembled prompt, references, an approved example, a don't list, an owner and a last-tested date.
  • Run a regression set whenever a model updates, and adjust snippets before client work changes.
  • Align the AI style system with existing brand guidelines.

Try it

Write four modular prompt snippets for a brand and assemble three prompts for different image types. Start a do/don't gallery with at least two approved and two rejected examples with reasons.