Google Ads & Meta Ads AI agents: 50 Claude roles, one for every job in the account.
Running Google + Meta ads is fifty small jobs done weekly. Most of them are checklists, not judgment — so we turned every checklist into an agent. Each one is a setup guide, not a loose prompt: open Claude, connect your ad accounts, paste the role.
Built by our community of 2,000 marketers
The full 50-agent ads library, free
Setup guides for Claude, ChatGPT and Perplexity.
Clients we work withClients we work with
Ryze
AI autopilot for paid ads
✓+25% traffic and DA 40+, or your money back
✓Rewrites competitors’ best pages onto your site
✓100+ high quality backlinks per month
✓Gets your items ready for ChatGPT Instant Checkout
✓Matches your site to customer demand from 39 sources
Every ad account breaks down into the same fifty small jobs: search terms burning spend, creatives fatiguing, budgets drifting, tracking breaking, the Monday summary. Done weekly, they compound. Skipped for a month, ROAS quietly erodes.
None of them need judgment — they need a checklist, a data source, and a weekly run. That's an agent. So we wrote all fifty as Claude agents:
Every agent is a setup guide, not a loose prompt — paste once, re-run with two words.
First pass ships real work: a kill list, a negatives list, a refresh queue — 20 minutes to first finding.
9 categories, 50 jobs — the whole ad account, Google and Meta.
Manage Google + Meta ads from Claude — the real setup
Ad assets auditor — sitelinks, callouts, images — coverage and performance
Landing page QA — speed, message match, broken forms before spend hits
Impression share tracker — where you lose auctions: budget or rank
Tracking & data · 4 agents
Conversion tracking health — fires, dedups, matches backend truth
Attribution comparer — platform numbers vs GA4 vs reality
UTM hygiene — one taxonomy, no more (not set)
GA4 reconciler — explains why platforms and GA4 disagree
Intelligence & reporting · 5 agents
Reporting — the Monday summary: spend, ROAS, what moved and why
Spend anomaly detective — catches the 3am budget spike before it hurts
Competitor ad spy — new competitor angles from the ad libraries, weekly
Forecast model — projects spend and revenue from current pacing
Ads roadmap — turns every agent’s findings into a prioritized plan
The core six
The six agents doing the heaviest lifting
These six produce most of the ad account's weekly movement. Each entry is the complete setup guide: what the agent reads, what it ships, and the full role — paste it into Claude and it's live.
Every ad account has campaigns that should have been killed weeks ago and winners starving on capped budgets. The auditor pulls 7 days of Google + Meta, scores every campaign against breakeven ROAS, frequency and CTR trend, and hands you one table: kill, scale, or fix — ranked by dollars at stake.
ReadsGoogle Ads + Meta via MCP — 7 days of spend, ROAS, CTR, frequency by campaign
ShipsKill / scale / fix table ranked by monthly spend at stake, one recommended action per row
CadenceEvery Monday — last week’s table is stale by Wednesday
ROLE PROMPT — PASTE INTO CLAUDE
You are my Ads Auditor.
Job: every Monday, tell me exactly which campaigns to kill, scale or fix.
Process:
1. Pull the last 7 days from Google Ads and Meta: spend, ROAS, CTR, frequency by campaign. Compare vs the prior 7.
2. KILL: ROAS below breakeven for 2+ weeks with meaningful spend.
3. SCALE: ROAS above target, impression share or budget capped.
4. FIX: falling CTR, rising frequency, or rising CPA — name the likely cause.
5. Ignore campaigns under $50/week — noise.
Output: one table — campaign, platform, spend, ROAS, verdict, the action. Ranked by spend at stake. Then a one-line total: how much weekly spend the kill list frees up.
02Flags rising frequency and falling CTR before ROAS craters
Creative fatigue agent
Creatives die quietly: frequency creeps past 3, CTR slides 20%, and ROAS follows a week later. By the time the dashboard makes it obvious, you’ve paid for the lesson. The agent watches frequency and CTR trend per ad, flags fatigue while the creative is still profitable, and keeps a refresh queue ranked by spend on the dying ad.
ReadsMeta + Google video/display ads — frequency, CTR, ROAS by ad, 14 days trailing
ShipsFatigue watchlist: ad, frequency, CTR trend, spend at risk, and a refresh brief per flagged creative
CadenceTwice a week — fatigue moves faster than weekly reporting
ROLE PROMPT — PASTE INTO CLAUDE
You are my Creative Fatigue agent.
Job: catch dying creatives while they are still profitable.
Process:
1. Pull 14 days by ad: frequency, CTR, ROAS, spend. Split into current 7 vs prior 7.
2. Flag ads where frequency > 3 AND CTR fell 20%+ week over week.
3. For each flagged ad, note spend at risk (current weekly spend) and days until ROAS hits breakeven at the current decay rate.
4. For the top 5, write a one-line refresh brief: keep the angle, change the hook / format / opening frame.
Output: the fatigue watchlist ranked by spend at risk, then the refresh queue. No ad under $30/week.
03Finds queries burning spend, ships the negatives list
Search-term miner
Search terms reports are where money leaks one irrelevant click at a time — and where nobody looks weekly. The miner reads every query that spent money, splits converters from burners, and ships two lists: negatives ready to paste, and converting queries you should promote to exact-match keywords.
ReadsGoogle Ads search terms report via MCP — 14 days, all campaigns
ShipsNegatives list grouped by theme (paste-ready), plus converting queries to add as keywords
CadenceWeekly — leaks compound daily
ROLE PROMPT — PASTE INTO CLAUDE
You are my Search-Term Miner.
Job: every week, stop the queries that burn money and promote the ones that make it.
Process:
1. Pull 14 days of search terms with spend, clicks, conversions.
2. BURNERS: spend > $10, zero conversions, intent clearly wrong. Group them into themes (jobs, free, DIY, competitors…).
3. KEEPERS: converting queries not yet added as keywords — recommend match type and campaign.
4. Check the burners against existing negative lists — only ship what’s actually missing.
Output: negatives grouped by theme, paste-ready; then the keeper list with target campaign. One line on total wasted spend found.
04Pacing vs plan — drift caught daily, not at month-end
Budget guard agent
Budgets don’t blow up at month-end, they drift there — a platform “learning phase” here, an unnoticed bid change there. The guard checks month-to-date spend against plan per account and platform, projects the month’s landing point at the current run rate, and proposes reallocations from losers to winners while there’s still month left to act.
ReadsGoogle Ads + Meta spend by day and campaign, plus the monthly budget plan you give it
ShipsPacing report: MTD vs plan, projected month-end, and a reallocation proposal ready to approve
CadenceDaily at 9am — drift caught on day 8 is a tweak, on day 25 a write-off
ROLE PROMPT — PASTE INTO CLAUDE
You are my Budget Guard.
Job: keep every account on plan, every day.
Process:
1. Pull month-to-date spend by account, platform and campaign. Compare against the monthly plan I give you.
2. Project month-end spend at the trailing-3-day run rate.
3. Flag anything pacing >10% over or under plan, with the driver (CPM shift, budget change, new campaign).
4. Propose reallocations: from campaigns pacing over with below-target ROAS, to campaigns capped with above-target ROAS. Amounts, not vibes.
Output: one pacing table (account, MTD, plan, projection, delta) and the reallocation proposal. Flag anything needing action TODAY at the top.
05Channel split, search themes and listing groups, exposed
PMax analyzer
Performance Max is a black box by design: one number out, no channel detail in the UI. The analyzer pulls what the API actually exposes — channel split, search themes, listing group performance — and tells you whether PMax is finding new demand or just cannibalizing your brand and shopping traffic.
ReadsGoogle Ads API via MCP — PMax channel data, search themes, listing groups, brand overlap
ShipsPMax reality report: where the money actually went, cannibalization estimate, listing groups to fix
CadenceEvery two weeks — enough data to see through the smoothing
ROLE PROMPT — PASTE INTO CLAUDE
You are my PMax Analyzer.
Job: tell me what Performance Max is actually doing with the budget.
Process:
1. Pull PMax campaigns: spend and conversions split by channel (shopping, search, video, display) as far as the API exposes.
2. Estimate brand cannibalization: overlap between PMax search terms and my brand terms.
3. Listing groups: flag products eating spend with no conversions, and starved products that convert.
4. Compare PMax ROAS vs my standard shopping and search campaigns.
Output: the reality report — channel split, cannibalization estimate, top listing-group fixes, and one verdict: feed it, trim it, or restructure it.
06The Monday summary: spend, ROAS, what moved and why
Reporting agent
Ad reporting eats hours and answers “what are the numbers” instead of the only question that matters: what changed, and why. This agent writes the Monday summary with a cause attached to every swing — it knows what the other agents shipped, what the platforms changed, and what competitors launched.
ReadsGoogle + Meta week-over-week via MCP, plus the change log of what shipped
ShipsOne-page Monday summary: spend, blended ROAS, movers with causes, and what ships this week
CadenceEvery Monday, before standup
ROLE PROMPT — PASTE INTO CLAUDE
You are my Reporting agent.
Job: every Monday, write the ads summary a smart stakeholder actually reads.
Process:
1. Pull last 7 days vs prior 7: spend, ROAS, CPA by platform and campaign.
2. Top 5 winners and losers by profit change — not spend change.
3. Attach a cause to every mover: check the change log I give you (what shipped, when) and platform changes. If unclear, say "unclear" — never invent one.
4. Lead with one headline: total spend, blended ROAS, direction, the single most important thing that happened.
Output: max one page. Headline → winners (why) → losers (why, next action, which agent owns it) → what ships this week. Plain language.
How to build one
What separates an agent from a prompt
A prompt is a one-off request. An agent is a standing role with four parts — take any prompt you already use and add the missing three:
01
Role
A written job description, not a question. "You are my Ads Auditor. Job: every Monday…" — the same words every run, so the behavior never drifts.
02
Data
A live connection, not pasted screenshots. Google + Meta via MCP means Claude pulls spend, ROAS and search terms itself. This is the step that turns a prompt into an agent.
03
Output spec
A fixed deliverable defined in the role: the exact table, the exact columns, ranked by the same metric. Comparable week over week — that's what makes the findings trustworthy.
04
Cadence
A schedule, not a whim. Save the role as a Claude project or skill and re-run it with two words: run weekly. The loop is the agent.
Every setup guide in this library is those four parts pre-written. That's the whole trick — and it's why the six roles above aren't "just prompts": each one names its data source, its deliverable, and its schedule.
Proof, not promises
What the first pass ships
One ad account, one afternoon. Accounts connected, six agents pasted in, first pass run end to end:
KILL LIST
campaigns below breakeven, ranked by spend at stake — ready to pause
NEGATIVES
wasted search terms grouped by theme — a paste-ready negatives list
20min
from setup to first finding — account connect included
Or skip the setup — your ads, on autopilot.
✓Runs this whole 50-agent stack for you, 24/7
✓Audits, creatives and budget shifts applied automatically
A prompt is a one-off request. An agent is a standing role: a defined job, a named data source, a fixed deliverable, and a cadence. You paste the role into Claude once, connect the data (usually Search Console via MCP), and re-run it weekly with two words. The output format never drifts, so results are comparable week over week.
Do I need to know how to code to set these up?
No. Every agent in the library is a paste-in setup guide. The only technical step is connecting Claude to Google Search Console through an MCP connector, which is a guided click-through — no terminal required. Setup to first finding took 20 minutes in our live run.
What access do the agents get?
Read access first: spend, ROAS, CTR, search terms, audiences. Nothing changes in your accounts until you approve a specific action — then Claude applies it through the same connector. It’s standard OAuth; you can revoke it any time from your Google or Meta settings.
Are the 50 agents really free?
Yes. The full library of setup guides is free — every role prompt, data checklist, and output spec. You need your own Claude subscription to run them. If you’d rather not run 50 agents yourself, Ryze runs the equivalent stack as a managed ads service.
Which agent should I set up first?
The Ads Auditor. One run gives you the kill / scale / fix table across both platforms — the fastest path from setup to money saved, which is why it’s agent #1 in the library.
Will Claude change my campaigns on its own?
No. The agents read data and produce recommendations — a kill list, a negatives list, a reallocation proposal. Changes are applied only after you approve them. If you do want fully hands-off management, that’s what Ryze’s managed ads service does on top of this library.
Ahrefs
Shopify
