How Atlas actually makes ads
Alex's description of the real process, and what the system must therefore do. This supersedes any earlier assumption that a good creative is an approved angle plus a layout.
Status: design. Very little of this is built.
The diagnosis that produced this document
Reviewing the first real output, four complaints, all with one cause:
- the proof card and the offer card were near-identical, differing only in visual design
- the native text post was three words on a white background
- the spotlight headline was "a staple, yes, but really very, very basic"
- nothing knew *when* a given format was the right choice
None of these are format or layout failures. The angle and copy objects carry almost no information — a one-line hook and a one-line promise. There is nothing in them for a format to express differently, so different UI over the same thin input yields the same ad twice.
Two ads that differ only in visual treatment also violate the creative-diversity principle the agency measures clients against. Diversity has to come from the idea, not the frame.
The floor is the content model, not the canvas.
The worked example that defines "insight"
Alex's own, and the clearest statement of what the system must be able to derive:
Inspire ID is CaseTrust accredited. The renovation industry has a lot of scams. CaseTrust is a good way to counter that. But renovation is not something people do all the time. So "Are you constantly sick of getting scammed in your renovation?" does not make sense — they are not always renovating. Instead: *"Hearing a lot of renovation scams in Singapore..."*
Read what that required:
| Input | Source |
|---|---|
| CaseTrust accreditation | brand knowledge base — exists |
| Renovation has a scam problem | industry research — does not exist |
| Accreditation counters that specific fear | synthesis — does not exist |
| Renovation is infrequent, so the fear is ambient not personal | purchase-context reasoning — does not exist |
| "in Singapore" | localisation — does not exist |
One of five inputs exists. That is the whole gap, and it is why current output is "very generic and very weakly worded". The difference between the wrong line and the right one is not phrasing — it is knowing that renovation is a once-a-decade purchase.
What has to exist
1. Industry research, per client. Comprehensive, and done well rather than gestured at: biggest pain points, challenges, common solutions, relevant institutions and accreditations, what people actually say on the ground (Reddit and similar), and how any of it localises to Singapore. Possibly driven by a questionnaire the system researches against itself.
2. Brand intelligence, deepened. The tab exists; the depth does not.
3. An orchestrator that makes these talk to each other. This is the load-bearing piece. Industry research alone is a report. Brand facts alone are a list. The insight lives in the combination, and nothing currently combines them.
4. A richer angle model. Not one-liners. An angle needs to carry the insight, the audience problem in context, the purchase frequency, which verified fact supports it, the objection it answers, and the awareness level it is pitched at — so that a format has something to work with.
5. A format database with selection reasoning. Formats grow and the system should learn them, but the important half is *choosing*: which format suits this insight, for this client, given what has worked before and which points are strongest. Chosen on semantics, not surface features — otherwise it is slop.
6. Historical performance. Which messages worked for this client, which did not, what budget remains, and therefore what to run. This is the AI media buyer's recommendation, and it should surface in the generator as "recommended for this client".
The visual principle
"I do not want statics that are purely posters where every client looks the same, always a template. What I want is a templated format, but the visuals — the way it brings out the points and the insights — can be different."
So: the structure is templated and reusable. The expression is specific to the insight. A scene prompt should be derived from what the ad is arguing, not chosen from a shelf.
Images come from two places: Higgsfield, and Drive for real client assets. Drive is deferred; the concept of the ad is the current problem.
Scope
Ads first, deliberately. The tooling built here — research, synthesis, angle modelling, format selection — is what later serves other content types. Ads are the sharp end and the place the quality gap is most visible.