Guardrails that keep a brand safe, written once so nothing embarrassing ships

Generative production makes volume cheap, which means the number of assets a client sees goes up, which means the number of chances for something wrong to reach them goes up with it.

The failures that matter are not quality failures. They are category failures: an image that contradicts the brand's positioning, includes something they will not be associated with, or breaks a rule they told you about once, months ago, in a call.

Nobody catches these reliably by looking. They are caught by having written them down and checking against the list — which is boring, mechanical, and the entire defence.

• Never depict — subjects, settings or activities the brand will not appear alongside.
• Never claim — words implying a property they cannot substantiate. This is frequently a legal matter, not a taste one.
• Never show — competitor products, other brands' marks, anything trademarked that is not theirs.
• Always include — required marks, required disclaimers, minimum clear space around a logo.
• Format rules — where the logo may sit, what may overlap it, permitted colourways.

The first three are the ones that cause damage. The last two cause revisions.

You will not guess a brand's guardrails accurately at the start. You learn them from what gets pushed back, and the discipline is to write each one down the moment it arrives.

Distinguish two kinds of rejection. 'We do not like this' is taste and belongs in art direction notes. 'We cannot show this' is a rule and belongs in the guardrail list. The second kind is permanent and applies to every future asset.

Ask directly at the start, too. A short list of questions in the kickoff — what can you not be shown near, what claims are off limits, whose marks must never appear — gets you most of the list before the first delivery instead of after the third.

Put the list into the brand file and have the delivery step check against it. Where a rule is machine-checkable — required ratios, asset counts, presence of a required mark — enforce it automatically and refuse to package a set that fails.

Where it is not machine-checkable — whether an image reads as being near a forbidden category — make it a printed checklist that is worked through before sending, item by item, with a signature. Not a mental review.

The value is that the check is the same on a calm Tuesday and on a rushed Friday. Human vigilance is not; that is exactly when things get through.

Anything on the never-show list should also be an explicit exclusion in the prompt's constraint block. Models add logos, text and background objects readily; naming them as forbidden removes most occurrences before review.

No logos or brand marks of any kind. No readable text.
No other products in frame. No third-party packaging.

This does not replace the check — models ignore instructions sometimes — but it changes the check from finding many violations to confirming there are none, which is a much more reliable task to perform.

A brand can get images anywhere. What is scarce is someone who will not put them in a position they have to explain internally.

Showing a client your guardrail list — that you wrote down what they told you and built a check around it — does more for the relationship than any amount of craft. It is the difference between a supplier and someone they trust with the account.

It also protects you. When something does go wrong, a documented process that was followed is a completely different conversation from an apology.