What gets lost when you turn your best designer into an agent
A Japanese creative-AI practitioner points out that the receiver of a studio's codified knowledge used to be a junior, and is now an agent that does the work. The result can be a degraded copy of your best people, at scale, with nobody asking whose knowledge it was.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page

CreativeEdge CL+, a Japanese practitioner who makes independent films with generative tools, published episode 125 of their radio series on 15 November, and the episode notes contain one observation I have not seen made this plainly in English.
The summary in my own words: for as long as organisations have written things down, the person receiving the written-down knowledge has been a junior. Manuals and know-how collections were, in the end, tools for growing people. A veteran's judgment became explicit, travelled through a manual or a training course, and landed in a newcomer. The veteran kept a place in the organisation as the person who teaches and the person trusted with the hard calls.
When the receiver becomes an agent, that chain changes shape. Explicit knowledge now goes straight into the thing that executes the work, and it does not pass through a person on the way.
The part of the job that was never in the manual
The notes are blunt about what a naive version of this produces. The unglamorous adjustments, the ones that read the room and cannot be justified in a sentence, get shaved off during the conversion to data. What is left is described as a degraded copy of a good employee, and the warning is that you can produce those at volume.
That matches what I see in studios building prompt libraries and style presets. The parts that survive documentation are the parts that were already easy to say:
- The label has to stay legible at a 1200 pixel export, not just in the preview.
- No rim light from camera left when the key is camera right.
- Keep the horizon below the eyeline for this account, above it for that one.
Those are real rules, and they are worth writing down. But they are not where the taste lives. The taste lives in the veto that arrives before the reason: this is competent and it is wrong for this client. Not this frame, not this week, not for this product. Ask the art director why and you get a sentence that sounds arbitrary until you have seen the three previous campaigns.
A rule set built from the first list will produce work that passes every stated check and misses the brief. It is the difference between a generator that reproduces the average of a studio's past output and a designer who knows which of those past decisions no longer applies.
Documentation is a ceiling, not a floor
It is tempting to read this as an argument against building the thing. I do not think it is. An agent that knows your legible-label rule and your horizon convention is genuinely useful, because those are twenty decisions you no longer have to repeat. The mistake is expecting the agent's ceiling to be your taste when the ceiling is actually your documentation.
Two consequences follow, and they are both operational.
The first is that the exceptions file is more valuable than the rules file. Every time a rule gets overridden, write down what overrode it and what the situation was. After twenty exceptions you have something a rule set cannot give you: the shape of the boundary. It is also the only raw material an agent can learn from, because the boundary is exactly what your senior staff spend their time defending.
The second is that the person who can write the exceptions is the person whose job feels threatened by the agent. Which brings up the harder half of the episode.
It also explains why the studios that survive model churn best are the ones holding assets that do not belong to a model. The churn playbook applies here: a decision log and a set of reference frames outlive whatever model you are prompting this year, and they are the only part of the operation you can move to the next one.
The incentive nobody designed
The notes put it directly: companies want the tacit knowledge of their strongest people, and the strongest people are wary of being duplicated. That is not paranoia. If the whole value of a senior is the judgment nobody else can reproduce, then handing that judgment to a system that reproduces it is a negotiation about their future, not a data-entry task.
CreativeEdge's suggested answer is to position senior staff as the maintainers of what the agent knows. Their word for it is roughly caretaker: the person responsible for tending the knowledge, which lets an organisation say with a straight face that this is not a replacement.
I would go further on the practical side, because "maintainer" is a title until it has hours attached. If curating the decision log is a real responsibility, it needs to be in someone's week, with a name next to it and a reason for them to care. The studios that get this right will look like the ones that already treat documentation as craft rather than admin: a senior spends a fixed slice of the month on it and is evaluated on whether the agent got better as a result.
Whose knowledge is it
The last third of the episode notes is a list of questions rather than answers, and they are the ones a studio should actually sit with.
An employee's professional ability is not entirely acquired at the company. Some of it predates the job, some of it came from a previous studio, some of it is the residue of a personal obsession with a particular kind of light. Treating all of it as company property is a stretch, and everyone involved knows it.
The list of things worth institutionalising, condensed: state the purpose, the scope, the permitted uses and the retention period of anything you collect. Explain what the agent does to roles and staffing before people find out by watching. Design the reward for contributing, in money, in evaluation, in career terms. Name the human accountable for the agent's decisions and the conditions under which an expert has to be pulled in. Collect only what the work needs. And separate the right to use knowledge from the right to the money it generates, because those are two different deals and most policies collapse them into one.
For a small studio, four of those fit on a page:
- A decision log that records overrides and exceptions, with the author's name on each entry.
- A stated rule that contributing to it counts as work, with hours and a credit.
- A named owner for escalations, so "the agent decided" never becomes an answer to a client.
- A written limit on what gets collected, so nobody has to guess where the line is.
None of that requires a legal team. All of it requires deciding, before the agent exists, that the people whose judgment you are encoding are parties to the arrangement rather than inputs to it.
The version I would actually build
I like the distillation, and I think the current wave of agents will quietly reorganise how studios hold knowledge whether or not anyone plans it. What I do not believe is that the resulting system will be better than the documentation behind it, or that the documentation will improve on its own.
The senior designer in this story is not a resource being extracted. They are the only person who can tell you which of your rules is wrong, and they will only do that if the arrangement gives them a reason to. An agent built from a studio's paperwork will behave like a studio's paperwork. If that is the ceiling you want, fine, but price it honestly, and do not call the result taste.