5.1AI visual guidelines for a brand
The essentials in 30 seconds
- A guideline, not stunts: aesthetic pillars + variable-based template prompts + locked parameters + shared references + prohibitions — so ten visuals from three people look like they came from the same hand.
- Built by induction: generate broadly, select what “is” the brand, analyse the constants, formalise the formula — never the other way round.
- Sustainable or nothing: an approach that is simple, constrained and reproducible; a style that is too complex cannot be replicated and the guideline dies.
Running case — Sonia, freelance graphic designer in Algiers
Sonia, a freelance graphic designer in Algiers, produces visuals for three small clients with AI — and everyone improvises their own prompt: her clients' feeds look like a patchwork. She decides to build AI visual guidelines for her main client: she generates broadly, keeps the ten visuals that “are” the brand, extracts the constants from them and formalises a variable-based template prompt. This case returns in the lesson, the diagram, the script and the 48h challenge — and continues in 5.2, 5.3, 5.4 and 5.5.
This same case returns in the lesson, the diagram, the script and the 48h challenge — and continues in 5.2, 5.3, 5.4 and 5.5.
1. What AI visual guidelines are
Producing one-off visuals is one thing; producing a coherent visual identity at scale is another. AI visual guidelines set how your brand generates its images — without them, everyone invents their own prompts and communication becomes a patchwork. Sonia sees it with her clients: three people, three styles, zero identity.
They contain: aesthetic pillars (medium, palette, lighting), standardised template prompts per use case, locked parameters (model version, stylize, formats), references (style image, style identifier, in-house LoRA) and prohibitions — including what must never be generated.
2. Building them by induction
You do not decree guidelines, you discover them. 1) Generate broadly, select what “is” your brand. 2) Analyse the constants: palette, medium, lighting, framing, level of detail, presence of people. 3) Formalise a template prompt formula with variables. 4) Add a shared style reference (or an in-house LoRA, section 4.3). 5) Document: template prompts, accepted AND rejected examples, usage and verification rules (sections 0.4 and 5.5). Sonia formalises it like this: “[subject] — flat vector illustration, limited palette of deep blue and warm sand, minimal details, soft even lighting — --ar 16:9” plus a shared style reference.
Guidelines must stay sustainable: a style that is too complex cannot be replicated — prefer a simple, constrained approach. That is what makes strong identities, with or without AI.
3. Keeping them alive with several hands
The real test starts when others generate: provide fill-in-the-blank templates per use case (each person only fills in the subject), share the assets (style references, version, parameters), show accepted and rejected visuals with the reason for rejection, and appoint a checkpoint who validates before publication. Require traceability: every visual accompanied by its prompt, version and parameters.
Treat the guidelines as a living document: models evolve and a new version can change your outputs — retest your template prompts with every major version. And remember the structuring rule: AI illustrates an idea, it does not document reality — products, premises and teams are photographed for real.
Did you know?
Guidelines without documented rejections are guidelines half written.
🔎 Concrete example — Sonia's template prompt: everything locked except the subject
Everything locked except the subject.
« [subject] — flat vector illustration, limited palette of deep blue and warm sand,
minimal details, soft even lighting — --ar 16:9 »
→ + a shared style reference (--sref).Reading: everyone only fills in the subject variable; medium, palette, lighting, format and parameters are imposed.
Lesson: consistent visuals even with several hands, with accepted AND rejected examples documented.
Section vocabulary
- AI visual guidelines
- The document that sets how the brand generates its images to stay consistent across several hands.
Simple example: Sonia writes pillars, template prompts, parameters and prohibitions for her client.
What it is not: a list of adjectives with no variables and no checks.
Reflex: do ten visuals from three people look like they came from the same hand?
- Template prompt
- A standardised formula with variables: everyone only fills in the subject, everything else is imposed.
Simple example: Sonia locks medium, palette and lighting; the client only changes the subject.
What it is not: a prompt reinvented every time by everyone.
Reflex: is the variable the only thing to fill in?
- Locked parameters
- The model version, stylize and formats locked to avoid drift.
Simple example: In Algiers, Sonia imposes the version and the --ar per medium.
What it is not: settings changed by feel with every generation.
Reflex: are the version and formats written down in black and white?
- Brand reference
- The shared asset embodying the style: style image, identifier or in-house LoRA.
Simple example: Sonia shares a common style reference with the whole client team.
What it is not: a text description that everyone interprets their own way.
Reflex: does the shared asset guarantee more than text?
- Rejected examples
- Non-compliant visuals documented with the reason for rejection; highly instructive.
Simple example: Sonia shows three rejected visuals: too saturated, off-palette, misleading photorealism.
What it is not: abstract rules with no examples at all.
Reflex: do we also show what must not be done?
In practice — Write your AI visual guidelines
- Generate broadly, then select 5 to 10 visuals that truly represent your brand.
- Analyse their constants: medium, palette, lighting, framing, level of detail, presence of people.
- Formalise a template prompt formula with variables + the locked parameters (version, stylize, --ar per medium).
- Add a shared style reference, then document: template prompts, accepted AND rejected examples, verification rules.
Ready-to-say script — proposing guidelines to the client
“To keep your visuals consistent even when produced by several people, I suggest AI visual guidelines: I generate broadly, together we select the ten visuals that really are your brand, and I formalise a variable-based template prompt with locked parameters and a shared style reference. I will also document accepted and rejected examples, with validation before publication — and together we will check the usage rights of each visual before release.”
To say to the client to get the guidelines approach approved, before talking tools or pricing.
Copy-paste template — AI visual guidelines sheet (1 page)
AI VISUAL GUIDELINES — 1-page sheet
BRAND: ___ · USE CASE: ___ (article / social / banner)
PILLARS: medium ___ · palette ___ · lighting ___ · framing ___
TEMPLATE PROMPT: « [subject] — ___ — --ar ___ »
LOCKED PARAMETERS: version ___ · stylize ___ · formats ___
REFERENCES: style ___ · LoRA ___
PROHIBITIONS: ___ (e.g. photorealism of real people)
VALIDATION: approved by ___ before publication · retested on ___Solo / small budget
Sonia alone: one template prompt and one style reference are enough for one client — the guidelines fit on one page.
Small business / small team
Three accepted and two rejected visuals pinned on the wall: the whole team understands the guidelines through examples.
SME in Algiers
In Algiers, Sonia sets a single validation checkpoint before publication for all the client's visuals.
Starting team
Pair: one generates and proposes, the other accepts or rejects with reasons — the example library grows fast.
✓ Best practices
- Build the guidelines by induction (generate → select → analyse), not from theory.
- Favour a simple, constrained, reproducible approach, with rejected examples.
- Treat the guidelines as a living document: retest template prompts with every major version.
✗ Common pitfalls
- A style too complex to replicate from one image to the next.
- Letting everyone invent their own prompts: the patchwork that dilutes the brand.
- Forgetting rejected examples and the validation checkpoint before publication.
Key takeaways
- AI visual guidelines ensure visuals produced by different people look like they came from the same hand.
- Contents: aesthetic pillars, template prompts with variables, locked parameters, references (style/LoRA), prohibitions.
- Built by induction: generate → select what “is” the brand → analyse the constants → formalise.
- Favour a SIMPLE, constrained approach: a style that is too complex cannot be replicated.
48h challenge — apply tomorrow
Fill in the guidelines sheet (template above) for 1 real use case of Sonia: does the template prompt fit in one sentence with variables?
- I can explain AI visual guidelines in one sentence.
- My variable-based template prompt is written, with locked parameters.
- I know why the guidelines are retested with every model version.
Frequently asked questions
How do I guarantee consistency when several people generate visuals?
Combine fill-in-the-blank template prompts (everyone only fills in the subject), shared assets (style references, version, parameters), a library of accepted AND rejected examples, a checkpoint before publication, and traceability (prompt, version, parameters attached to every visual). Treat the guidelines as a living document, retested with every major version.
Should everything be AI-generated, or mixed with traditional visuals?
Mixing is usually preferable if deliberate and consistent: AI illustrates an idea, it does not document reality — products, premises and teams are photographed for real. Unify both sources (palette, shared grading, framing) and plan transparency about generated visuals (section 5.5).
What are AI visual guidelines for?