Higgsfield as the art director

Revises the recommendation in visual-system.md. That document argued for compositing in code and treated "let the image model render the whole ad" as tempting and wrong. After probing what the MCP surface actually exposes, that position does not survive.

What changed my mind

Alex's WANTED poster settles the design question on its own — aged paper, illustrated subjects, typographic hierarchy, stamps, a value stack read as a reward table. satori cannot produce that and never will. It is not a layout I failed to write; it is a class of output a CSS-subset renderer does not reach.

Then the MCP surface turned out to hold considerably more than raw generation:

Surface	What it is
42 ad formats	Visual layout archetypes with preview media — Then vs Now, Comparison Table, Magazine Style, Bold Statement, Whiteboard Explainer, Organic Post, Benefits Checklist
26 video presets	Direct-to-Camera, Selfie Testimonial, Secret Hack Reveal, Before and After, TV Spot, Wild Card
~60 viral presets	Distinctive visual treatments — Noir, Magazine, Ink Riot, Paper, Vintage, Hand Paint, Sticker Peel
`brand_kit`	Brand name, hex colours with usage, fonts, logo, tone of voice, aesthetics, industry
`ad_reference`	Upload a real ad; it extracts the concept

The last two remove the two objections I raised:

*"Brand colours are approximate and the logo is not really the logo"* — there is a brand kit entity taking hex values with declared usage, plus a logo URL. Push the client's kit in and generation is brand-aware.
*"We would have to build the taste ourselves"* — ~130 professionally tuned treatments already exist. Authoring that library is many months of design work we are not going to do.

Resourcefulness was the standing instruction and this is the clearest case of it in the project.

The one real objection, and how it is answered

The genuine problem is not quality. It is that our guardrail controls what a creative says, and if Higgsfield renders the words, that control is gone. Approved copy could arrive on the page altered, and the whole approval chain would become decorative.

The answer is to read the render back:

approved copy ──▶ Higgsfield renders ──▶ OCR / vision read-back
                                              │
                                              ▼
                              does rendered text match approved copy?
                              does it show a third-party brand mark?
                              does it contradict the knowledge base?
                                              │
                                    reject ◀──┴──▶ accept

This is the visual critic from visual-system.md doing double duty, and it converts "let the model render everything" from a guardrail hole into a closed loop. Verification after generation replaces control during generation — which is the only workable answer when the renderer is not ours.

It also, finally, closes register #55: the generated hinge carrying a competitor's mark would have been caught by exactly this read-back.

Division of labour

Atlas owns	Higgsfield owns
What is true — brand and market knowledge bases	How it looks — presets, ad formats, Marketing Studio
What we are arguing — insight, angle, copy	Rendering, at a quality we cannot reach in code
Whether it is allowed — the approval chain	Video generation, clipping, analysis
Whether it is good — taste rules, critics
Verification of what came back

Restated: Atlas is the strategist and the editor. Higgsfield is the studio. We never lose the argument or the approval; we stop pretending we can out-design a purpose-built tool.

Two catalogues, two axes

Our 42-format catalogue and Higgsfield's 42 ad formats overlap in name and differ in kind, and the difference is the useful part:

Ours describes the argument — what the ad claims, at which awareness level, and critically failsWhen. Higgsfield has no equivalent; it does not know that a problem-led format is wrong for a once-a-decade purchase.
Theirs describes the render — the visual archetype, with a preview and a working generator.

So they compose rather than compete: ours selects the argument, theirs renders it. The mapping between them is a real artefact to build, and unmapped formats on either side are interesting — one of ours with no render is a gap, one of theirs with no argument is an opportunity.

Viral presets sit on a third axis again: visual *treatment*, orthogonal to both. Format decides the structure, preset decides the world it lives in. That is precisely "templated format, visuals specific to the insight".

What the agent actually does

Concept — from the insight, decide what this ad should show and why it argues this point
Select — our format for the argument, their ad format for the render, a preset for treatment
Brief — approved copy, brand kit, concept, constraints, into the Higgsfield call
Render — Marketing Studio or generation
Read back — OCR and vision check against approved copy, brand marks, knowledge base
Accept or reject — and per the taste rules, rejecting everything is a legal outcome

Steps 1, 2, 5 and 6 are ours and are where the judgment lives. Steps 3 and 4 are theirs.

Still to study

Higgsfield Academy. Alex flagged it and he is right — a purpose-built studio publishing its own teaching material is the fastest route to understanding what its tools are actually good at, and what conventions its presets assume. Not yet read.