How much brand context do you give an agent before it becomes useful?

Examples and boundaries tend to travel better than adjectives. Give the system approved work, rejected work with reasons, audience evidence, claims it may make and claims it must never make. Then test whether it can explain the difference before asking it to create.

Rejected examples helped us more than the brand deck. Especially when we added one sentence explaining why each example was wrong.

Customer interview snippets. The model finally stopped making up polished problems and started using the words buyers actually use.

I would not give it everything at once. We made a small context pack for one audience and one offer, then expanded it after the first few reviews.