16 Meta ads use cases · newest Nov 12, 2026 · all 87 →
AI agent use cases for Meta (Facebook and Instagram) ads.
Real Meta (Facebook and Instagram) ads workflows people are running with AI agents, newest first. Each one links to the original post, lists the tools used, and includes the prompt where it was shared.
Nick Stebbings (Agentiff.AI) describes a six-agent setup for Meta ad accounts: an Analyst finds wasted spend and performance changes, a Watchdog catches CPA spikes and falling ROAS, a Fatigue Hunter spots creatives losing steam, a Strategist pulls winning patterns and recommends tests, a Creator builds new static ad variations, and a Reporter writes client reports and account summaries. The output is one list of what needs attention, what to fix, what to test and what to make next, and the team still reviews and approves every decision. A useful template for splitting a long account review into small, checkable jobs.
@tobiadonadon_ described a V2 skill that runs inside Claude Code or Codex and every morning reports which ads are wasting money, which are selling and which are finished. Nothing changes without your approval. It cuts ads on the numbers rather than one bad day, leaves new ads alone until they've had a fair run, writes a fresh opening for ads people are tired of seeing, enforces a hard spend cap, flags settings Meta quietly switches on, and logs every approval with its date. The flagging step matters as Meta turns on more defaults (see the Marketing API v27 Product showcase and device defaults). He gives V1 away and charges for V2.
Every morning, review my Meta ads and label each one: wasting money, selling, or finished. Don't judge ads that haven't had a fair run yet, write a new opening for any ad with fatigue, list any settings Meta turned on by default, and wait for my approval before changing anything. Log each approval with the date.
@williamkast_ shared a five-step agent workflow: connect Claude to your Meta Ads MCP, run deep research (review mining, comment mining, Reddit), analyze the last 90 days of ads, categorize each winner by angle and persona, then compare that map to the research so you see which personas, angles and awareness levels never ran. He reports accounts often scale on only 1–3 core angles, and filling the gaps has driven 2–3× growth within a few months. Useful for brands that want the next creative lane from their own winners plus outside voice-of-customer, not more lookalike statics.
Connect to this Meta ads account. Analyze the last 90 days of ads, categorize each winner by angle and persona, then mine reviews, comments and Reddit for missing personas, angles and awareness levels. List the gaps and propose the next 5 ads to test.
@ZedNilm1 published a free Claude skill for ecommerce Meta accounts that reads purchases, ROAS, CPA, AOV and MER, flags ads that spent 2× target CPA with zero sales, spots creative fatigue (frequency over 3.5 or CTR down 30%), proposes scale steps capped at +20% per change, and enforces a daily spending limit it cannot pass. Every OK is logged, and nothing moves a dollar until you approve. It can also pull long-running niche ads from gethookd to clone. Useful for brands that want an agent on Meta without unsupervised spend.
Audit this Meta ads account for ROAS, CPA, creative fatigue and zero-sale spenders. Propose scale or kill options with a max +20% budget change, stay under the daily spending limit, and wait for my approval before changing anything.
In Claude, add Meta's official ads MCP connector (mcp.facebook.com/ads), connect Slack for reports, and create a project with your product info, ideal customers and competitors. Paste instructions that make Claude your media buyer: every day review Meta ads, propose new creatives, pause underperformers and scale winners, never spend without approval, and summarize in Slack. Put it on a daily scheduled task so you only approve from your phone. The quoted Spanish thread claims one brand 10x'd revenue in a month with the same pattern.
You're my media buyer. Every day, review my Meta ads, find what's working, propose new creatives, pause the underperformers and scale the winners. Never spend anything without my approval. Summarize everything in Slack.
Tobi Adonadon shared a free Claude Code skill for founders who run their own Meta lead-gen ads. Twice a day it reads spend, views, leads and cost per lead, a leads file showing who replied and who's serious, competitors' live ads in the Ad Library and a web search for the sector's typical cost per lead. It then proposes a few numbered moves, each with the number behind it (for example "Pause A2. $9 spent, 0 leads" or "A1 budget +20%: $10 to $12 a day"), and you reply with the ones to do. Nothing runs without your yes, and four hard limits sit on top: an account spending limit, a daily cap, at most +20% per change, and a log of every approval. Ads with no hook, no leads after three times the target cost, or that Meta stops delivering are switched off with a replacement ready. He reports cost per lead going from $3.15 to $1.84 in 4 weeks on an example account (self-reported) and says it suits small and mid accounts, not huge ones.
Ilkhom (@sandiegocausa), a Shopify and Amazon FBA seller, used Claude Code to build a responder and moderator for the comments on his Facebook ads. It hides damaging comments, answers shoppers' questions about the product and delivery (attaching product photos and links when useful), and replies to happy buyers asking them to share pictures. He says it runs on its own once built, without Claude Code in the loop, and replaced the paid comment-moderation apps he'd otherwise use. He didn't share the code; treat it as a pattern: give the bot your product facts, shipping policy and a list of what to hide versus answer, and review its replies before Black Friday volume hits.
Khush Mahajan (@YesKhush_5) built a Meta Ads skill and gave it $1,500 a day across several products. It has four jobs: build an identical account skeleton for each new product (campaigns, ad sets, audiences, exclusions, naming, conversion events verified before anything spends); upload ads from a creative sheet as drafts only, checking every ad has media, thumbnail, primary text, headline and CTA, and never publishing until he types publish; analyze hook rate against hold rate and cost per result per ad, funnel drop-off and placement/age/gender against a fixed exclusion bar; and give every ad one verdict (scale, watch or kill) with the number behind it and a review date. An 'ad brain' spreadsheet it reads at the start and writes at the end of each session holds config and unit economics, a change log (old value, new value, why, what should move, when to review), changes parked because they'd reset learning, ad rotations, learnings and open questions. Each proposed change is sorted into safe now, resets learning (batch it), or waits on a running test. It may not report a change as done until it has read it back from the live account and matched the activity log. He says it had been running three days when he posted.
Build each piece in Claude Code first: a script that pulls your ad numbers into a daily sheet, the instructions the bot follows on every run, and a rules file listing what it may never touch. Then give a scheduled agent one job. Every morning it reads the sheet (not the ad account) and flags ads where frequency is rising while click rate falls, and ad sets that spent with zero leads. It can only suggest changes. Every bot reads and adds to one shared file about what you sell, who buys and what you've already tested. Any change in the ad account waits for approval, and the spending cap is set inside the ad platform, so no instruction can push spend past it. When it keeps making the same mistake, open its last 10 runs in Claude Code and fix the step that causes it.
read this bot's instructions and its last 10 runs
find the step that causes the mistake i marked
propose one change and explain why
wait for my approval before editing any files
Claude Code reads every image in a creatives folder, writes 3 headline variants for each and uploads them to one ad set as paused ads through the Marketing API (about 50 ads in a couple of minutes, per the author). You then switch on the ones you like in Ads Manager. To keep the pipeline from breaking every 60 days, it uses a Meta system user token set to never expire (employee role, with only the page, ad account, app and Instagram assets it needs) kept in a .env file, not a Graph API Explorer token.
Pulls 30 days of ad-level data, picks the three ads with the lowest CPL, and writes 20 new hooks and angles on the same offer and landing page. They're created in a new ad set, all paused, and the agent is told never to activate or change budgets.
goal: 20 new ad variants for ad account [ID]. use the meta ads MCP to pull the last 30 days at ad level and pick the 3 ads w/ the lowest CPL. write 20 variants (new hooks & angles, same offer, same landing page). create them under a new ad set, all PAUSED. never call ads_activate_entity and never change budgets. when you're done, send me the list: ad name, hook, which winner it's based on.
Pulls last month's numbers from Meta and the client's analytics, rebuilds the campaign around what worked, refines the creatives, and sets everything up through the Meta Ads MCP with cookieless tracking.
Load a claude.md with your ICP, offers, voice and a log of tests you've killed. Then a Monday audit pulls 30 and 90-day performance, top and bottom ad sets, creative fatigue and pixel health, and outputs five prioritized actions for the week.
A /meta-ads-audit skill that pulls live campaigns, ad sets and ads, flags frequency over 3, under-pacing below 70% and over-pacing above 110% of budget, and checks audience and pixel health.
A Claude Code subagent that scores a Meta account PASS, WARNING or FAIL across 46 checks, covering Pixel and CAPI health, event match quality, creative diversity and fatigue, learning phase and Advantage+ setup, and lists quick wins under 15 minutes.
A local MCP server with 24 tools and skills for /ads-launch, /ads-report, /ads-optimize and /ads-pause. A brand folder keeps voice and defaults consistent, and new campaigns are always created paused.