AI Ad Pipeline (the orchestrator)

The connective skill for an AI advertising campaign. It does not write a brief, direct a video, make a still, or QA a frame. It sequences the skills that do, owns the handoffs between them, and holds the one mandatory human gate. It is the advertising peer of /planning-pipeline (which chains the plan reviews) and /ship-it (which self-drives a build): a skill whose whole job is to run other skills in the right order.

Use it when the owner wants a whole campaign built or planned, for any product. The machine is product-agnostic; only the brief changes.

Two laws sit above everything: never start with generation, start with the brief; and machines make many, humans plus the ad algorithm pick few. Generate wide, tag everything, gate hard at one human checkpoint.

First Rule

Do not start with generation. Start with the product file, the argument, the audience, the compliance lane, and the measurement plan. Every promoted product has a PRODUCT-BRIEF.md (what it is, what changed, allowed claims, forbidden claims, audience, naming, visual rules, verification checklist) before any script, prompt, image, clip, or distribution plan. If any compliance source is missing, stop before creative. For current platform rules, prices, and disclosure requirements, run /last30days (or live web research in Codex) at use time; treat all of them as drift-prone.

The pipeline (chain the subskills, in order)

Fix the CONCEPT and the GENERATION METHOD before you build or run QA (Codex R2/R5). The corrected order: a reference is analyzed, the concept passes the concept-fit + producibility gates (in vid-ad), THEN a storyboard, THEN binary criteria are written and approved, THEN production, THEN auto-QA against those criteria, THEN the human gates. Building QA around a concept that fails the gates, or a mechanic the engine cannot make, is effort spent on the wrong target.