From a chat history to a managed canon

Making imagery with a paid AI subscription works, right up to the point where the number of images grows and a second person joins. Then five problems surface, none of which are solved by prompting better.

Five problems

1. The model drifts. Despite an extensive context, instructions are followed only partly and the model fills in the rest. That is known behaviour, not an operating error.
2. The knowledge sits somewhere nobody can reach. A canon living in a personal chat history is not searchable, not version controlled, not transferable and not auditable. Lose the account and you lose the visual identity.
3. There is no route to another model. Switching means explaining everything again, so there is in effect no freedom of choice, not even when another model is demonstrably better for a specific job.
4. There is no version control on images. Feedback goes into the chat as a loose sentence and the result is a new file with no provable relation to the previous one.
5. Only one person can work with it. Every extra colleague would have to learn the whole style again and would feed the AI their own, differing prompts.

What is put in their place

The application stores the canon as data, sends it along fully and automatically, checks the result, keeps everything with versions and provenance, and lets several people and agents work side by side without the style drifting apart. Read on at the canon, the library and quality control.

What is deliberately left out

• Building the app or website the imagery is made for. The Art Director delivers material to it.
• Final graphic finishing at prepress level: print preparation and colour separations stay a separate step.
• Personal data of end users. The application does not touch any of it; that is a hard boundary.
• Audio production, bookkeeping, invoicing and project planning.