a human media buyer works 8 hours a day.
my AI media buyer works 24/7.
while i sleep, my agentic Claude Opus 4.8 loop is checking our creatives every 15 minutes through the Meta’s MCP.
it automatically scales winners / pauses losers so I NEVER lose money from poor management, ad fatigue or forgetting to increase bids
so i documented the ENTIRE playbook on how you can do the same…
here’s what’s included inside the guide:
1. setup - Meta's MCP connected to Claude:
(system user token + ads_management perms, JSON config block, verify connection. Claude OAuth or Claude Code config for multi-account)
2. the master operating prompt:
(monitoring cadence every 15 min at peak, decision authority, reporting format with a live account health score)
3. the decision rules:
(creative fatigue at CPM up 20% w/w, CTR down 15%, thumbstop under 25%, frequency over 3.0)
3. the 15-minute ROAS shift detection prompt:
(trailing 3-hour vs 7-day ROAS, real-time reallocation capped at 15%)
4. hook performance segmentation:
(sorted by naming convention - AUTH, CONT, PAIN, PROOF - to see which angle carries the spend)
5. creative rotation workflow:
(min 10 approved creatives tagged BACKUP, rotated like-for-like, plus the rotation prompt)
6. what it doesn't replace:
(strategy, offer, funnel logic still come from you. a force multiplier on a good strategy, not a fix for a broken one)
all backed by $10M+ generated for service-based businesses through AI meta ads funnels
and for 24h you can access this entire system for FREE
like + comment "MEDIA" and I’ll send it over.
(must be following + RT for priority access)
after generating $10M+ for clients' service based businesses with meta ads funnels...
i've realized that the #1 thing that kills scaling attempts isn't:
> the targeting
> the offer
> the media buyer
it's not testing ENOUGH ad formats
so i built an entire AI production engine that maps every Meta ad format to a specific prompt architecture (with opus 4.8), a specific visual model (inside Higgsfield MCP), and a specific funnel stage (inside the meta ads API)
here's everything that this system includes:
> the Higgsfield MCP setup
(the exact JSON config block, the file paths for mac and windows, and the one-line command to confirm it's actually connected before you generate a thing)
> the format assignment protocol
(one prompt that tags every brief across 4 points: awareness stage, placement, format family, angle, then routes it to the right format before a word of copy gets written)
> 10 format prompts, one per ad type
(founder selfie, talking head, B-roll w/ voiceover, split-screen, green screen, UGC, data viz, story-native, carousel, AI creative, each with its own word count, structure, and pacing rules baked in)
> the paired visual prompts for every format
(what to feed chatgpt images 2.0 for statics, what to feed seedance 2.0 for motion, exact dimensions per placement, all passed in through the MCP so you never leave Claude)
> which model runs which layer
(opus 4.8 for the script and reasoning, chatgpt images 2.0 for statics, seedance 2.0 for anything that moves, and what breaks when you mix them up)
> the 3-format testing matrix
(how we test every new angle across 3 formats and the exact CTR threshold at 200-300 impressions that decides what scales and what gets cut)
like + comment "FORMATS" and i'll send it over
(must be following + RT for priority access)