Skip to content

Google Ads AI Agentic System

Native Meta launch preparation and cross-channel evidence join the Google system.

Version
v1.4.0
from
v1.3.0
Released
Type
Feature

The September release added a guarded Meta lane, asset census, creative production tools, product-collection offers and Taboola audience preparation.

What to inspect first.

Suggested first inspection · based on this dated release
Your next stepWork and review
Inspect
Inspect one asset-census record before choosing a replacement.
Retain if you try it
Keep the asset identity, observed window, evidence floor and proposed swap together.
Review before use
Separate the census observation from a launch decision; check the source and permissions before using the Meta or audience-sync lanes.

Source: The Google Ads AI Agentic System v1.4.0 · dated published release · Reading guidance from the published release; no exercise, generated artifact or customer result is claimed here.

Use the instructions and compatibility guidance for the version and access you actually have. Open my system materials →

What changed

  • Native Meta Ads launch lane (4 new skills)

    meta-prep, meta-launch, meta-report, and meta-autogrow. Build a research-grounded campaign package, launch it paused behind a two-step human approval, read canonical performance, and run a weekly kill, refresh, and scale cycle. Meta now has the same guarded-launch discipline as Google.

  • Cross-platform asset census (1 new skill)

    brand-asset-census reads every live Google Ads asset and every Taboola item, judges each one in context with a spend floor and Poisson significance, and writes one Google Sheet per brand with a swap plan. Read-only, so it is safe to run any time.

  • Higgsfield creative studio (8 new skills)

    generate images, video, 3D, and audio, plus brand kits, product photoshoots, marketplace cards, explainer videos, YouTube thumbnails, consistent characters, and full websites. Ships with its bundled helper scripts.

  • Pollection offer pages (1 new skill)

    offer-pollection builds a Shopify page with a feed-approved buy box on top and a filterable category grid beneath it, the right destination for generic category searches where the buyer is still choosing between products.

  • Taboola first-party audiences (1 new skill)

    taboola-audience-sync onboards your buyer ledger as Taboola audiences and runs lookalike discovery, dry-run by default.

  • Leaner roster

    eight alias skills and two unused router skills were retired. Every capability they pointed at is still here under its canonical name.

  • The complete engine now ships, not just what each skill points at

    the full utility layer (987 Python utilities, up from 682), every core library, schema, data contract, page block, and seed data file is included, along with the scripts and templates that live inside skill folders. Nothing a skill calls is missing from your copy.

Release notes

This release rounds out the three-surface engine from v1.3.0 with a fourth native lane and a measurement layer that spans all of them. The End-to-End Codebase moves from 173 to 178 skills: 15 added, 10 retired, 450+ existing files deepened.

The headline: Meta Ads, launched natively (4 skills)

Until now the Meta lane repurposed winning Google creative and pulled performance. It now builds and runs Meta campaigns from scratch with the same gates as the Google engine:

  • meta-prep - prepares a campaign package for one deployed presell or offer destination: research spine, an entity-diverse creative slate, five true-angle primary texts, and specs with structure defaults. Everything is staged locally and paused. Cold traffic never routes to a product page.
  • meta-launch - the only skill that creates Meta campaign state. Gate verification, a validate-only preview, human approval, bottom-up creation in a paused state with read-back, then deliberate activation as a second approval. It hard-blocks when there is no verified purchase signal, so no signal means no spend.
  • meta-report - read-only campaign, ad set, ad, daily, and breakdown insights with canonical single-action metrics, attribution-window stamping, and spend-significance framing before any judgement.
  • meta-autogrow - the weekly optimization cycle: a read-only sweep produces a dated plan (kill candidates, fatigue refresh flags, scale unlocks), then approved changes execute one by one. Budget moves in 20% steps outside the learning phase.

Know what is actually working: brand-asset-census

One read-only observer over every live asset you run: Performance Max, Search, Demand Gen, extensions, feed product ids, and Taboola items split into title, thumbnail, description, and CTA. Each asset is judged in context over 90, 180, and 365 days under a per-platform spend floor with Poisson significance, stitched across campaign types, and written to one Google Sheet per brand alongside a swap plan. Run it with no brand to cover the whole portfolio.

Higgsfield creative studio (8 skills)

A full generative creative toolkit, included under its MIT license (see THIRD-PARTY-NOTICES.md in your repo):

  • higgsfield-generate - images, video, 3D assets, and audio, including image-to-video, restyling, and reframing.
  • higgsfield-brandkit - complete visual brand systems: palettes, logo marks, typography, mockups, packaging, and editable brandbooks.
  • higgsfield-product-photoshoot and higgsfield-marketplace-cards - product imagery and marketplace-ready listing cards.
  • higgsfield-video-explainer and higgsfield-youtube-thumbnail - explainer videos and thumbnails.
  • higgsfield-soul-id - consistent characters across generations.
  • higgsfield-websites - generated sites, with the bundled 3D asset pipeline scripts.

You need your own Higgsfield account to use these; nothing else in the system depends on them.

Pollection offer pages

offer-pollection adds a new page type to the offer pipeline. The top of the page is a feed-approved buy box (live price, rating, add-to-cart) so the URL stays valid for Shopping and Performance Max. Beneath it sits a curated, filterable category grid and a cross-sell rail. Use it as the bottom-of-funnel destination for generic category queries when you have six or more product groups in one family. It stages locally and the existing offer-deploy-shopify skill pushes it live.

Taboola first-party audiences

taboola-audience-sync is the Taboola sibling of brand-audience-sync on the shared buyer ledger. It onboards your buyers as audiences (one-time and repeat), runs lookalike discovery and seed matching, and generates the request for your Taboola rep, since lookalike creation is not available through the API. It mutates only when you tell it to and is a dry run by default.

A leaner roster

Ten skills were retired. None of them carried unique capability:

  • Eight alias skills - demandgen-create-image, demandgen-create-video, demandgen-upload-image, demandgen-upload-video, feed-report, search-ad-report, spy-demand-brief, and spy-product-brief. Each was a second name for a canonical skill that is still here.
  • Two unused routers - pipeline-run and workflow, along with their runner scripts. The gate chain inside each skill replaced them.

If you had a saved prompt that used one of these names, use the plural or canonical form instead (for example demandgen-prep-images, demandgen-upload-images, brand-report, spy-brief). SKILL-ROUTING.md maps every task to its skill.

Deeper engine across the board

More than 450 existing files were updated in this release: tighter gates in the search, Performance Max, and presell pipelines, new shared authoring standards for copy, creative, page copy, and reporting, the upload gate, new example campaign specs for Meta and Taboola, and new image and render creative templates.

The complete engine ships

Earlier releases included the supporting code each skill referenced by name. That missed modules that are loaded indirectly, so some utilities, libraries, and page blocks were absent from your copy. This release ships the system directories whole:

  • Utilities: 987 Python utilities, up from 682, including the full Google Ads and Microsoft Ads operation set, the presell page builder, and the quality-gate judges.
  • Core libraries, schemas, and contracts: 70 libraries, 116 schemas, and 29 data contracts, up from 55, 68, and 12.
  • Page blocks: all 87 presell and offer building blocks (hero, content, conversion, trust, interactive, and structured-data blocks), up from 19.
  • Skill support files: helper scripts, agent configs, and page templates that live inside a skill folder.
  • Deploy engines: the Shopify page deploy and PostHog patch engines that the deploy skills import.

Every file still passes the same sanitization gate, and an import check now confirms that every module the code loads is present.

By the numbers

  • End-to-End Codebase: 173 → 178 skills (15 added, 10 retired)
  • Native launch lanes: Google Ads, Microsoft Ads, Taboola, and Meta Ads
  • Utility layer: 682 → 987 Python utilities
  • Reference layer: 2,000+ skill instruction and reference documents
  • Blueprint curriculum: unchanged

Existing owners pull the latest from your private codebase repo - no upgrade charge. Pricing is unchanged.

Review current system scope