Guide: Automate Your Weekly Meta Ads Creative Loop with Claude + Higgsfield

Running Meta Ads at scale means constantly producing fresh creatives, monitoring performance, pausing what's not working, and scaling what is. For most sellers and marketing teams, the creative production bottleneck is what kills testing velocity. You know you should be launching more variations, but the overhead of briefing, generating, uploading, and organizing ads across campaigns makes weekly iteration unrealistic.

This workflow uses two MCP connectors inside Claude — Higgsfield for image and video generation, and Meta's official Ads MCP for campaign management — to build a repeatable Monday loop. Claude pulls your performance data, identifies winning patterns, generates new creative briefs and assets, uploads them to Meta, and produces a structured weekly report. You control how much runs automatically and how much requires your review.

The result is a system that handles the operational weight of creative production and media buying while you stay focused on strategy and sign-off.

This guide walks through the full setup and the first three steps of the weekly workflow. The complete SOP covers all six steps, the control database structure, deployment phases, and the Claude agent system prompt.

What You'll Need

MCP Connectors

• Higgsfield MCP (https://mcp.higgsfield.ai/mcp) — generates images and video using 30+ models including , Veo 3.1, Kling 3.0, Flux, and Seedance. Authenticates via OAuth with no API key required. Free tier includes 150 credits per month.

• Meta Ads MCP (https://mcp.facebook.com/ads) — the official Meta connector with 29 tools covering campaign creation, ad management, creative upload, performance insights, and diagnostics. Free during beta. No Developer App or API token required, just standard Meta Business OAuth.

Control Database

A Google Sheet or Airtable with four tabs that serves as the single source of truth the automation reads from and writes to: rules (budget caps, thresholds, launch permissions), creative_briefs (every concept tracked from idea through launch), performance (weekly metrics from Meta), and competitor_patterns (reusable structural insights from competitor research).

Account Access

• Meta Business Manager with ad account permissions

• Active Meta Pixel or Conversions API setup

• At least one product with existing ad performance data to analyze

Step 1: Connect Higgsfield and Meta Ads MCP

Start by adding both connectors in Claude.

For Higgsfield: Go to Settings → Connectors → Add custom connector. Name it Higgsfield, paste the URL https://mcp.higgsfield.ai/mcp, and authenticate.

For Meta Ads: Same path — Settings → Connectors → Add custom connector. Name it Meta Ads, paste the URL https://mcp.facebook.com/ads, authenticate, and select your Business Manager and ad account.

Test both connections before proceeding. For Higgsfield, try this prompt:

Use Higgsfield to generate a 9:16 product lifestyle image for [PRODUCT]. Direct-response style, clear product visibility, native social feel.

Confirm you get back an accessible asset URL.

For Meta Ads, test with:

List all Meta ad accounts available. For each, show account name, ID, and whether campaign creation is enabled.

If the official Meta MCP isn't available for your account, there's a fallback option using Pipeboard's Meta Ads MCP at https://meta-ads.mcp.pipeboard.co/.

Step 2: Set Up Your Control Database

Create your four-tab database before running the first cycle. The rules tab governs every automated decision — set these once, then adjust as you learn.

Key rules to configure: maximum daily budget per campaign, maximum total new campaign budget, minimum spend before any decision gets made, your target CPA and ROAS thresholds, CTR benchmarks, the percentage for budget scaling increases, spend limits before pausing non-converters, the frequency number that triggers creative refresh, allowed campaign objectives, and whether campaigns launch paused or active.

Two critical toggles control the level of automation. Set full_auto_launch_enabled to FALSE and full_auto_scaling_enabled to FALSE for your first few weeks. This means Claude does all the work — researches, creates briefs, generates assets, uploads, builds campaigns — but everything stays paused until you review and activate it manually.

The creative_briefs tab tracks every concept from idea through launch with fields for the creative ID, product, audience, angle, hook, format, visual concept, Higgsfield prompt, ad copy, brand constraints, the source pattern it's based on, status, asset URLs, and all Meta campaign IDs.

Step 3: Run Performance Analysis and Classify Ads

This is where the Monday loop begins. Claude pulls Meta performance data for the last 7, 14, and 30 days at account, campaign, ad set, and ad level, then classifies every ad using a decision framework.

The classification logic works like this: if CPA is below target and ROAS is above target, it's a winner. If CPA is good but frequency has crossed your threshold, it's a fatigued winner — still performing but running out of steam. If CTR is strong but CPA is poor, the hook is working but the funnel isn't converting. If spend has passed the pause threshold with zero conversions, it's an underperformer and gets paused. If CTR is declining more than 20% week-over-week, it's a fatigue risk. And if spend is still below the decision threshold, it simply needs more data.

Every classification gets logged to the performance tab along with the specific metric that triggered the decision. This creates the foundation for Steps 4 through 6 — generating new creative briefs from winning patterns, producing assets with Higgsfield, uploading to Meta, scaling winners, and pausing losers.

The rest of the SOP — including the creative briefing workflow, Higgsfield asset generation prompts, Meta campaign structure and naming conventions, the scaling and pausing rules, the full Monday report template, deployment phases, and the Claude agent system prompt — you can find here, [a free gift from me :)]

• Meta lost 20M daily active users in Q1 2026, but still reaches 3.56B people daily. Is Meta dying?

• Google killed FAQ rich results on May 7. You can strip the structured data from your code, but leaving it won't hurt. Other search engines may still use it.

• Google is adding "Further Exploration" links and "Expert Advice" snippets to AI Overviews, giving publishers more chances to appear in AI search results.

• Affirm and Klarna are now available as payment options inside Google's Gemini app and AI search mode, expanding BNPL into AI-powered shopping.

• Higgsfield launched a Virality Predictor. Upload any clip under 15s and get a viral score, hook rating, hold rate, and a heatmap of which brain regions your video activates.

⚡ Visual Electric - An image generator built for the creative process.

🪄 Magic Documents: Organize, rename, summarize, and extract key data from any document with AI.

🖱️ Persuva - Craft ads that click and convert.

📅 Howie.ai - Say goodbye to scheduling hassles with advanced AI coordination

🤖 Gobble bot - Digest your website, PDF, YouTube Video, Notion pages and more into one text file to upload to your GPT chatbot (link)

😍It was great, all I needed

🙂Good, not great

🙄It sucked