The Creative Director: A Hundred Variants and One Idea

By Tuesday of week six Theo Alvarez has two hundred image variants and no campaign. The variants took an afternoon and cost about the price of lunch. Several are genuinely beautiful. Laid out together on a wall they look like the output of eleven different companies, and a plant operations director in Ohio who saw four of them across four months would not connect any of them to each other.

That is the whole problem of this module, and it is not a problem agents caused — it is one they made impossible to ignore. Choice used to be expensive and was therefore rationed. Now it is free, and the thing that was always scarce is suddenly the only thing that matters.

What lands on Theo’s desk

Camille’s messaging brief from module 5 gives him the line — you hear it from us, not from your customer — and, less obviously, a set of prohibitions that govern pictures. A crystal ball, a forecast curve arching into next quarter, a dashboard showing a problem already solved: each says predicts without using the word, and Wes refused that verb in module 2. An image is not exempt from a claim because it contains no sentence.

Ryan’s plan from module 6 gives him formats, quantities and dates, including the one that constrains everything: the booth graphics go to a printer in week eleven, three weeks before the trade show, which makes the creative deadline earlier than the campaign deadline. Naomi’s copy and claims sheet arrive from module 7. And behind all of it sits Joel’s finding from module 4, carried through Camille: the buyer’s real fear is being blamed for a shortage nobody saw coming.

What a creative director actually does

A creative director’s output is a decision, not an artefact. They choose the single idea a campaign expresses and then defend it against forty reasonable requests to make one asset slightly different. The defending is most of the job, because every individual exception is sensible and the sum of them is incoherence.

The software on Theo’s desk

Figma is where the idea is decided and tested — in one colour, at thumbnail size, at forty feet on a booth wall, and without the logo. No tool supplies that judgement; it supplies the surface on which the judgement is made.

The image agent earns its place in two specific ways: exploring a dozen directions in an afternoon before anything is committed, and then adapting the chosen idea across dozens of placements, which is the work that used to consume junior designers’ weeks. The asset library exists because a campaign with assets scattered across drives is a campaign where a withdrawn image survives somewhere. And the provenance log is the unglamorous file that makes module 9 survivable: every reference input, its licence, its origin. “The model made it” is not a chain of title.

The four decisions

Why volume is not creative direction

An image agent produces a hundred variants in an hour for a rounding error. It is easy to mistake that for a creative capability, and it is worth being exact about why it is not.

A campaign works when a plant operations director sees a LinkedIn advert in March, opens an email in April and walks past a booth in May, and knows without reading that these are the same thing. That recognition is built by repetition of one idea, not by quality of individual images. Eleven excellent images with nothing in common produce no recognition at all — the audience simply sees eleven unrelated adverts and remembers none of them. A hundred variants without an idea is a hundred ways of being forgettable.

There is a second, less obvious cost. A hundred options do not make the choice easier; they make it impossible to make well. Nobody can hold a hundred images in mind, so selection collapses into preference, and preference is settled by whoever is most senior in the room. What looks like a rich set of options is usually a decision about to be made badly.

So Theo generates before he decides, and then he throws almost all of it away. The two hundred variants were not a waste — they were an afternoon of thinking made visible. But they are input to a decision, not a substitute for one.

The idea he chooses

The idea does not come from the tool, and it could not have. It comes from two sentences written by other people.

Camille’s line promises lead time: you hear it from us, weeks before it reaches your line. Joel’s interviews say what that is worth — the fear is not an abstraction called supply chain risk, it is the Monday meeting where a plant director is asked why the line stopped and has no answer. Put those together and the campaign is not about visibility, or data, or supply chains. It is about the interval between the first sign of trouble and the moment it hurts you, and about who owns that interval.

So Theo’s idea is a gap. A small amber mark, a measured span of empty space, and then a heavy mark where the event lands. That is all of it. In an advert the gap is the layout, with Camille’s line set at the early mark. On a landing page it is the rhythm of the sections. In an email it is a single rule of type. On the booth wall it is forty feet of deliberately empty wall with the small mark near the entrance and the heavy one by the door, so a visitor physically walks the lead time before anyone speaks to them.

It also stays inside Wes’s refusal, which Theo checks deliberately rather than assuming. The gap depicts elapsed time between a signal and a consequence — something the product actually produces. It is not a curve arcing into the future, which would depict a prediction and would have been a claim nobody could support.

Then he tests it, and the tests are the part worth learning, because they are what separates an idea from a nice picture. Does it survive one colour? The gap is spatial rather than chromatic, so it does; trade press still prints mono, and half the campaign’s life is in places nobody art-directed. Does it read at thumbnail? Two marks and a space are legible at a hundred pixels in a mobile feed, which is the size most people will ever see it. Does it hold at forty feet on a booth wall? It gets better, because scale gives the empty space physical meaning. Does it work without the logo? After six weeks of repetition, the interval itself becomes the identifier — which is the real test, because it asks whether you own anything.

His other strong candidate fails those tests, and he kills it. A grid of nine icons for the nine risk signals looked superb at full size in a presentation, turned to grey mush at thumbnail, and said nothing whatever without a caption. It was a good illustration. It was not an idea.

Where agents genuinely help him

Having chosen, Theo now needs that one idea rebuilt for about forty placements — LinkedIn single image and carousel, six display sizes, email headers, webinar slide masters, the one-pager, eight booth panels — and this is where the agents are unambiguously excellent.

Adaptation is the largest gain. Recomposing one idea across dozens of aspect ratios is mechanical, skilled and enormous; an agent does the rebuilds and Theo checks the crops, the type sizes and the ones where the empty space has been squeezed until the idea stopped working. What used to take a junior designer two weeks takes a day and a careful review.

Exploration is the subtler gain. Before committing, Theo took twelve directions to a rough visual state in an afternoon; the old economics allowed three in a week, which meant ideas were defended because they had been expensive rather than because they were right. The value here is in the killing, not the keeping.

And abstract material — backgrounds, textures, gradients, the visual furniture that depicts nothing and no one — is now generated in minutes rather than licensed from a stock library at a cost per image.

The honest consequence is a staffing one, and it deserves neither doom nor evasion. The resizing work was a real job at Cadence and it has largely gone. It was also how junior designers learned: you understand composition by rebuilding one layout forty-eight times and watching where it breaks. Theo does not backfill the junior who leaves in week eight, so the team goes from six to five. He changes what the remaining two juniors do — from week one they direct and critique the agents rather than executing, which is harder work needing judgement they used to acquire slowly. He is candid, in the conversation with his own director, that he does not yet know whether that produces art directors as good as the apprenticeship did. Fewer hands and more judgement is the direction of travel; where the judgement will now come from is an open question, and pretending otherwise would be dishonest.

What he refuses to generate

Theo writes one prohibition and applies it without exception: no generated depiction of a real customer’s premises, people or products.

The rule costs him immediately. In the second batch there is an image of an oven plant at dusk — the light on the loading bay, the shape of the roofline — and it is the strongest single asset in the set. It reads unmistakably as Calder Thermal, because Calder’s own website photographs were fed in as the visual reference. Theo kills it in week nine.

His reasoning is not legal language, and it is better for that. It is a photograph of a place that does not exist, pretending to be a place that does. Every viewer who knows Calder will read it as their factory, and Cadence cannot stand behind a single thing it shows — not the building, not the plant, not the implication that this is what the pilot looked like. Permission to be named as a reference customer is not permission to synthesise a picture of your premises. An image of a real place made by a machine is a claim, and nobody at Cadence can substantiate it.

For everything else that survives, he keeps a provenance record: the reference inputs with their source and licence, the tool and version, the direction given, the date, and the named person who directed it. He does this because module 9 will ask, and because the answer cannot be reconstructed afterwards — a generated image carries no memory of what went into it, and neither does the person who made it, three weeks and two hundred variants later.

And the killing does not quite work, which is the most useful part of the story. The Calder image had already been dropped into a landing page mock-up in week eight, before Theo withdrew it, so a copy reaches Miriam’s folder in week ten and she blocks it independently — as module 9 records. Theo counts that as his failure rather than hers. Withdrawing an asset is not the same as removing it, unless there is one place approved assets live and everything else is known to be a copy. Building that place is Dana’s job in module 10.

Where this goes wrong

The first is the hundred-variant review. Options are laid out, everyone points at a favourite, and the campaign’s organising idea is decided by taste rather than by argument — usually the taste of the most senior person present, who has not read the messaging brief. The remedy is not fewer options; it is presenting one idea with its reasoning and using the variants as evidence, not as a menu.

The second is subtler and increasingly common. Generated imagery converges: ask several models for a striking industrial photograph and they return the same lighting, the same lens, the same mood, because they were trained on the same material. A team producing a hundred variants can therefore produce a hundred images that look exactly like every competitor’s campaign. Distinctiveness has become scarcer precisely because volume became free, and the only defence is an idea that came from somewhere other than a model.

What Theo hands on

Miriam Adeyemi in module 9 receives the approved creative package with the provenance log attached, and that log is what converts her rights review from an investigation into a check. Where a claim in Naomi’s copy resolves to a source in minutes, an image without a record of its inputs resolves to nothing at all — so the assets she can clear fastest are the ones Theo documented at the moment they were made. She also receives the withdrawn-image note, and finds a copy of the withdrawn image anyway.

Dana Whitlock in module 10 receives the master files, the specifications for every placement and the constraint that matters most: the booth panels print in week eleven. Anything not approved by that date ships as it stands or does not ship. Review time and print time overlap by a week, which nobody planned and everybody now has to work around.

Designing this desk’s agent: the variant agents

Theo’s agents are adaptation machines, and their risk sits almost entirely in what goes in rather than what comes out.

One concrete design choice. Store provenance beside the asset rather than in someone’s memory. Every generated image carries its prompt, its reference inputs and their licences — because “the model made it” is not a chain of title, and module 9 will ask.

{
  "asset_id": "img-hero-04",
  "prompt_id": "p-882",
  "reference_inputs": [
    {"source": "brand_library/plant-abstract-02.jpg", "licence": "owned"},
    {"source": "stock/id-99412", "licence": "extended_commercial"}
  ],
  "depicts_real_entity": false,
  "model": "image-gen-v4",
  "approved_by": "theo.alvarez"
}

The metric to track. Provenance completeness, which has to be 100% rather than high — a single asset with unknown inputs is the one that causes the problem. Alongside it, the rejection rate at review, which tells you whether the constraints are actually reaching the people writing prompts.

One idea, or a hundred options?

Read each one and decide what Theo would do, then tap a card to check.