Xandrov StudioAtlas of shipped systems
← The territory/ II. Applied & Vertical AI/ № 06 — Ad Assembly Line
Dossier № 06 / 19 — Applied & Vertical AI — Applied AI · Creative pipeline

Ad Creative Assembly Line

Structured volume, not one clever headline.

File
№ 06 / 19
Territory
Applied & Vertical AI
Classification
Applied AI · Creative pipeline
Status
Working pipeline · Packaged offer
§ 01 Context

Direct-response teams live by testing velocity. They don't need one clever headline — they need structured volume: dozens of angles, hooks, headlines, body-text blocks, and complete ad variations, pushed out fast enough to keep pace with A/B testing cadence. Single-prompt LLM generation produces inconsistent output. Real creative work has stages: angle informs hook, hook informs headline, headline informs body, and all of it informs the final assembled variation. The pipeline is the product.

§ 02 The problem

Single-prompt LLM generation produces creative that lacks internal consistency. One call tries to handle too much — angle, hook, headline, body copy, even assembly — and the outputs inevitably compete and contradict each other. Effective creative work happens in discrete stages, each stage's output feeding into and constraining the next. Most LLM generation pipelines skip this structure entirely, treating creative as a one-shot problem.

§ 03 The approach

A Node and Express server accepts a product or offer brief and runs it through a staged generation chain using Gemini. The pipeline processes each stage in sequence: angles first (multiple strategic framings of the offer), then hooks (attention-getters that attach to each angle), headlines (the hook made clickable), primary text (body copy or CTA), and finally assembled variations (complete ads combining all preceding outputs).

Each stage streams back over server-sent events as it completes. The first usable asset arrives within seconds, not after the entire run finishes. Every stage remains inspectable — the team can review angle outputs before hooks are generated, pause if needed, and restart with adjustments.

The architecture means each prompt's job is narrow and repeatable. An angle prompt always takes the brief. A hook prompt always takes an angle. A headline prompt always takes a hook. This stage-to-stage constraint keeps outputs consistent and lets teams rebuild a single stage if results are weak — rather than re-running the entire pipeline.

Structured volume beats one clever line — and the pipeline is where speed comes from.

§ 04 Decisions that mattered
D1
Staged chain over single prompt
Each stage is inspectable and steerable rather than locked into one mega-prompt's output. A team can pause, review, and adjust before the next stage runs.
D2
Server-sent event streaming
First usable assets arrive in seconds, not after the full pipeline completes. Users see angles while hooks are still being generated.
D3
Narrow prompts per stage
Each prompt handles one job only — angles, then hooks, then headlines. The output of one stage becomes the input constraint for the next, maintaining consistency through the pipeline.
§ 05 Where it stands

The system is a working pipeline, offered by Xandrov Studio as a packaged service. It takes a brief and produces structured creative volume on demand — angles, hooks, headlines, body text, and complete ad variations — all streamed in real time.

Node + ExpressGeminiSSE streaming
§ 06 — Sanitization record

No specific client projects, user counts, testing metrics, or performance data are included. The dossier describes the technical architecture and the service offering itself; actual generated outputs remain proprietary.

Have a system like this worth building?

Fixed scope, quoted after one working session, first output inside two weeks. Read by a person within one working day.