Three ways to tell AI what on-brand means for your visual brand

To teach AI your visual brand guidelines, capture the visual judgment your reviewers already make and feed it back as shared guidance. Three signals do the work: edit diffs at the final approval gate, quick votes on each asset and votes on refresh candidates. Distilled together, they give every future generation the same rules for that organisation.

Fifty acts of visual judgment a week produce zero learning when every correction disappears inside an exported file.

The fix is not a longer brand prompt. It is a feedback loop that captures three kinds of judgment:

1. Lock-gate edit diffs capture what changed before approval.
2. Per-asset votes reveal which visual directions work repeatedly.
3. Refresh-candidate votes show where the brand can evolve.

Those signals distill into one shared guidance layer. That guidance is then injected into every future generation for that organisation.

After reading, you can run a three-lane visual review and turn everyday feedback into reusable visual brand guidance.

Lock-gate diffs turn final edits into brand rules

The lock gate is the final approval point before an asset becomes part of the organisation’s accepted work.

That moment contains unusually strong evidence. A reviewer has moved beyond preference and decided, “I am willing to put our name on this.”

An edit diff compares the generated candidate with the locked version. It captures specific changes such as:

• Replacing a blue studio background with warm daylight
• Cropping tightly around the product instead of showing a wide office
• Removing decorative icons from a chart
• Reducing headline size to give the image more space

The diff captures what changed. A short note captures why it changed.

For example:

Change the blue studio lighting to warm natural light. The cooler image makes the product feel clinical. Keep the close crop.

The system can distill that into reusable guidance:

Prefer warm natural light for product imagery. Avoid clinical blue lighting. Use close crops when the product is the focal point.

One approach that worked for us is asking for a reason only when the reviewer changes the visual direction. This keeps approval fast while preserving the judgment behind meaningful edits.

Per-asset votes reveal repeated visual preferences

A lock-gate diff provides deep evidence from one approved asset. Per-asset voting provides lighter evidence across a larger set.

A thumbs-up says, “More like this.” A thumbs-down says, “Do not repeat this direction.”

The vote becomes more useful when the reviewer selects the visual element that drove it:

• Composition
• Colour
• Lighting
• Subject
• Typography
• Image density

Consider a fractional CMO reviewing five campaign images for one client. Three staged office scenes receive negative votes. Two close product shots receive positive votes.

That 60 percent rejection rate on one subject type is stronger than one isolated comment. It gives the system a direction to test against future approvals.

A useful negative vote could read:

The staged meeting scene feels generic. Keep the off-centre composition, but show the operator using the product.

The resulting guidance can preserve the part that worked without repeating the rejected subject matter.

This distinction matters. “I dislike it” teaches very little. “Keep the composition and replace the staged subject” gives the next generation a usable constraint.

Refresh-candidate votes separate brand evolution from drift

Current brand guidance answers one question: what looks like us now?

Refresh-candidate voting answers another: where are we willing to go next?

Without that separate lane, experimental work can contaminate established guidance. One high-contrast campaign image might become an accidental rule for every future asset.

A refresh review works better when each candidate receives one of three decisions:

• Preserve: This belongs in the existing visual system.
• Explore: This direction deserves further testing.
• Reject: This moves away from the intended brand.

The reviewer then names what can change and what stays fixed.

For example:

Explore the higher contrast and tighter crop. Preserve the existing typeface and restrained colour palette.

That vote can become scoped guidance:

Campaign visuals may use higher contrast and tighter crops. Keep the current typography and core palette unchanged.

The scope is the important part. It lets the visual brand evolve through deliberate decisions instead of absorbing every experiment.

Usable visual feedback names the change, reason and scope

“Make it pop” creates another review round because neither a designer nor an AI system can locate the decision inside it.

Signal-rich visual feedback has four parts:

1. Element: Name the visible object under review.
2. Direction: State what to increase, reduce, keep or replace.
3. Reason: Connect the change to the intended perception.
4. Scope: Say where the rule applies.

Here is a complete example:

Reduce the number of background objects. The current scene makes the product feel harder to use. Apply this to product-launch images, but keep detailed environments for technical explainers.

This format gives the system an observable change, a business reason and a boundary.

It also protects useful exceptions. A sparse launch image and a detailed technical diagram can both be on-brand because they perform different jobs.

Shared guidance makes one correction useful to the whole team

The three lanes carry different weights.

A locked edit records final conviction. An asset vote identifies a repeated preference. A refresh vote proposes a future direction.

T2D3 OS distills those signals into shared visual guidance for that organisation. Every future generation reads the same guidance instead of relying on personal prompts or remembered comments.

The consequence compounds. One person’s five-second correction can improve the next asset generated by everyone in the same organisation.

For fractional CMOs, the organisation boundary also matters. Feedback for one client stays attached to that client’s visual system. It does not bleed into the next account during a busy four-client week.

Start with one complete visual review cycle

For the next visual batch, route final edits through the lock gate. Add quick votes to every candidate and review experimental directions in the refresh lane.

Use the four-part feedback format when a vote needs explanation: element, direction, reason and scope.

Then review the distilled guidance before the next generation. The practical test is simple: the next asset should reflect the previous decision without someone retyping it into another AI tool.