How to use loops in Claude Code for Google and Meta ads.
A loop is a goal plus a stop condition the agent runs on a schedule — not a prompt you re-type each day. Pointed at your accounts through MCP, it pulls yesterday’s Google and Meta numbers, flags what’s off target, and drafts the fixes. Here’s the setup, eight copy-paste loops, the verifier gate, and the cost guardrails.
Jump to the 8 loopsBuilt by our community of 2,000 marketers
Free Claude loops, skills and prompts
Copy-paste loops for Google & Meta, plus skills for Claude Code.
Clients we work withClients we
work with
Start here
What a loop actually is
Most people use an AI agent the slow way: ask, wait, read, fix, ask again — every turn runs through you, and the moment you stop, it stops. A loop changes who drives. You state the goal once and the agent runs the cycle itself, on a schedule. In Claude Code the command is /loop.
For ad management the five-part cycle maps straight onto a daily optimization routine:
The loop cycle
↻Runs on a schedule and repeats until the gate passes or the stop condition — success, or a maximum number of tries — is hit.
Verify is the part people get wrong. Without a real check on each pass, you do not have a loop — you have the agent agreeing with itself on repeat. The check has to be something that can fail the work: a CPA above target, a ROAS below floor, a spend cap breached. The model that proposed the change is far too generous grading its own homework, which is exactly why an unattended ad loop needs an objective gate and a human approval step before it moves money.
Why it fits ad management
Recurring, checkable work is exactly what loops are for
The same checks every morning, the same reallocation every week, the same search-term cleanup every few days — that is the shape a loop fits. A human reviews three to five data points well; one account surfaces hundreds of changed signals overnight. The loop reads them all, applies your rules consistently, and never skips a Monday.
| Task | Manual, by hand | As a loop | What you do |
|---|---|---|---|
| Daily spend & anomaly check | 20-40 min/day | Runs at 7am | Read the flagged list |
| Search-term → negatives | 1-2 hrs/week | Drafts a negatives list | Approve or trim |
| Budget reallocation | 2-3 hrs/week | Proposes shifts by ROAS | Approve within caps |
| Creative-fatigue review (Meta) | Often skipped | Flags rising frequency | Brief new creative |
The honest version: loops shine on tactical, repeatable work with an objective check. They do not replace strategy, brand judgment, or the decision to enter a new market — keep those manual. For where the line sits between AI execution and human strategy, see our guide to rules-based vs agent-based Facebook ads management.
Set it up once
Connect Google and Meta, then point a loop at them
A loop earns its setup cost only when four conditions hold. Miss one and a single well-aimed prompt is faster and cheaper than the machinery around it.
- 1The task repeats, at least weekly. Daily monitoring and weekly reallocation qualify. A one-time audit does not — just prompt it once.
- 2Something can automatically reject bad output. A ROAS floor, a target CPA, a spend cap, or a significance rule. No gate means the agent grades its own work.
- 3The agent can reach the data and the controls. Live API access to your accounts, not screenshots pasted in. More on the connection below.
- 4Your token budget can absorb the waste. Loops re-read context and retry whether or not a run ships a change. The technique scales with budget.
The connection is what makes any of this real. By itself, Claude can only read data you paste in and draft changes you copy back. To let a loop act, you connect Claude Code to your platforms through MCP connectors that call the official APIs:
How a loop reaches your accounts
Your ad accounts
MCP connectors
Official APIs
Scoped OAuth + dev token
Claude Code loop
Runs on a schedule
Discover → plan → verify
Approval gate
🔒 You approve
Then applied to the live account
By default the loop reads and drafts. Write access — changing budgets and bids — stays behind the approval gate until you trust it.
- •Google Ads through an MCP connector on the Google Ads API. See configuring MCP for multiple ad accounts.
- •Meta through the Marketing API and Graph API. See pulling Facebook ads data into Claude.
- •GA4 and Search Console for the conversion side, so the loop optimizes to real outcomes, not platform-reported clicks. See connecting every source to Claude via MCP.
Start every connection read-only while you test. Grant write access only once a loop has proven itself by hand, and keep it scoped to the accounts in your skill file.
The anatomy
The five building blocks of an ad loop
A working loop is assembled from a few pieces. A year ago this meant a pile of bash scripts only you could maintain; now the pieces ship inside Claude Code.
- •Automation (the heartbeat). What makes it a loop and not a one-off:
/loopfor a cadence, scheduled or cloud runs so it keeps going after you close the laptop, and hooks that fire at points in the run. - •Skills (project knowledge, once). A
SKILL.mdholding your account structure, target CPA/ROAS, brand rules, and the accounts the loop must never touch — read every run. - •Sub-agents (maker and checker). Split the agent that proposes a change from the one that reviews it. The proposer can be fast and cheap; the reviewer slow and strict. That separation is most of the quality.
- •Connectors (so it acts, not suggests). The MCP connections to Google, Meta, GA4 and Slack that let the loop apply a change and post the summary, instead of describing what it would do.
- •The verifier (the gate). The objective check — a ROAS floor, a CPA ceiling, a spend cap — that rejects a bad change automatically. The one piece that decides whether the loop helps or just spends.
One more piece sounds too simple to matter and is the spine of every loop that survives: a state file. A markdown file or shared board, outside the conversation, that records what the loop changed, what it is watching, and what it learned (“this campaign always dips on weekends, do not flag it”). The agent forgets between runs; the file does not.
Copy-paste
8 loops worth running for Google and Meta ads
Each loop lists the goal, a cadence, a prompt to adapt, and the gate that keeps it honest. Replace the bracketed values with your own targets, and keep budget and bid changes behind approval until each loop has earned your trust.
Loop 01 · Google + Meta · daily
Daily spend & anomaly watch
The single most useful loop. Every morning it compares yesterday against a rolling baseline and surfaces only what moved enough to matter, so you start the day reading a short flagged list instead of ten dashboards.
Gate: read-only. The loop proposes, never applies. Its only job is to escalate the right anomalies, so the “stop and report” instruction is the safeguard.
Loop 02 · Google · every 3 days
Search-term miner → negative keywords
Poor negative-keyword hygiene wastes 20-35% of Google Ads spend. This loop scans the search-terms report for queries that spend without converting and drafts negatives grouped by reason, so you approve a clean list instead of building it from scratch.
Gate: the conversion + CPA threshold is the objective check; nothing qualifies as a negative unless it clears the wasted-spend rule. Approval required before write.
Loop 03 · Google + Meta · weekly
Budget reallocation by ROAS
Money sitting in low-ROAS campaigns is the most common leak. This loop ranks campaigns by efficiency and proposes shifting budget from the weakest to the strongest — but only inside the caps you set, so no single move can blow up the account.
Gate: the significance minimum plus the per-campaign change cap and total budget cap. These hard limits are what make the loop safe to run unattended once approved.
Loop 04 · Meta · every 2 days
Creative-fatigue watch
Meta creative decays fast. When frequency climbs and CTR slips, costs rise before the dashboard makes it obvious. This loop catches fatigue early and briefs the next round of creative so the refresh is ready before performance craters.
Gate: the frequency-and-CTR rule. An ad only enters the brief queue when both signals cross the line, which keeps the loop from crying wolf on normal variance.
Don’t want to build and babysit loops?
- ✓Continuous loops on Google, Meta + 5 more
- ✓Built-in gates, approval and one-click rollback
- ✓No tokens to budget, no schedule to maintain
2,000+
Marketers
$500M+
Ad spend
23
Countries
Loop 05 · Google · daily
Target-CPA drift guard
Smart Bidding drifts as auctions shift. This loop watches for CPA creeping above target over a stable window and proposes a measured adjustment, rather than the knee-jerk daily edits that reset the learning phase.
Gate: the 5-consecutive-day rule plus the once-a-week limit. Both exist to stop the loop from over-correcting on noise and churning the bidding algorithm.
Loop 06 · Google + Meta · daily
Disapproval & tracking-health monitor
A disapproved ad or a broken conversion tag can quietly kill performance for days. This loop checks the unglamorous health signals every morning so a tracking break becomes a same-day alert instead of an end-of-month mystery.
Gate: read-only by design. Tracking and policy issues need human diagnosis, so this loop’s value is in catching them fast, not acting on them.
Loop 07 · Meta · weekly
Audience & placement exclusion miner
Advantage+ placements and broad audiences spread spend wide, and some lands on placements that never convert. This loop finds the consistent under-performers and proposes exclusions, recovering budget without narrowing reach more than the data justifies.
Gate: the spend + CPA + minimum-results threshold, plus a reach floor so the loop can’t over-prune. Approval required before any exclusion is applied.
Loop 08 · Google + Meta · weekly
Cross-channel performance digest
The reporting loop that saves the most meeting time. It rolls Google and Meta into one plain-language summary with the numbers that matter and a short recommendation list, written for whoever reads it — you, a client, or a CMO.
Gate: none needed — it only reads and writes a message. This is the safest loop to start with while you learn how the cadence feels.
Do it in order
Build the minimum viable loop
The order matters more than the tooling. Loops that fail in production almost always skipped a step. Build the smallest version first — one automation, one skill, one state file, one gate — and add complexity only when the simple version is reliable.
- 1Get one manual run reliable. Run the prompt by hand against your real account until the output is consistently right and safe.
- 2Turn it into a skill. Move the account context, targets and rules into a SKILL.md so the run is repeatable.
- 3Wrap it in a loop with the gate and stop condition. Add the objective check and a hard limit before it runs unattended.
- 4Then schedule it. Only once it is reliable by hand. Start read-only, keep spend changes behind approval, widen scope as trust grows.
Scheduling something you have not proven by hand is exactly how a loop quietly burns budget overnight. Prove it once, harden it, then automate it.
The catch
The risks and costs nobody mentions
Loops are powerful on the right task and expensive on the wrong one. Three problems get sharper as the loop gets better, not easier — and on ad accounts, two of them cost real money.
Loops fail quietly. An agent set to finish when “done” can decide it is done too early, exit on a half-finished job, and keep running and spending while producing nothing — engineer Geoffrey Huntley named this the “Ralph Wiggum loop.” On an ad account the equivalent is a loop that shifts budget on a flawed read and you only notice on the invoice. The fix is an objective gate that can fail the work, a hard spend cap, and an approval step before any change goes live.
The cost compounds. A loop re-reads its context on every pass, and that grows each run. A single medium task can run 50,000 to 200,000 tokens, and a maker-checker split doubles it. Track cost per accepted change, not tokens spent: if you discard more than half of what the loop proposes, you are doing the review work it was meant to save. Cap every loop with a token and iteration budget, and run cheap models on the boring steps.
Comprehension debt. The faster a loop ships changes you did not make, the larger the gap between what your account is doing and what you understand. Read the diffs. Spot-check the gate. The comfortable move — accepting whatever the loop returns without forming an opinion — is the dangerous one.
The throughline: build the loop, but stay the engineer. The loop changes the work; it does not remove you from it.
Build vs buy
DIY loops vs a managed platform
Building loops yourself gives full control and suits technical teams with the budget to run and maintain them. A managed platform like Ryze AI runs the same loop pattern as a product — it holds the API connections and applies approved changes 24/7, with the gates, approval and rollback built in. The difference worth remembering: Ryze AI executes approved changes, it does not just recommend them.
| Aspect | DIY loops in Claude Code | Ryze AI (managed) |
|---|---|---|
| Setup | MCP connectors, OAuth, dev tokens, skills | Connect accounts in ~15 minutes |
| Gates & approval | You design and maintain them | Built in, with change history + rollback |
| Cost model | Token usage you budget and cap | Flat subscription, no token math |
| Maintenance | Yours when an API or loop breaks | Handled for you |
| Best for | Engineers who want full control | Teams who want the outcome, not the upkeep |
Many teams hand-build one or two loops to learn the pattern, then move to a managed platform once they want loops on every channel without the upkeep. To compare the broader tooling, see our roundups of AI tools for Google Ads and AI tools for Meta ads in 2026.
Get a free instant audit
Get a free, instant read on your paid ads or SEO — and fix it right away.
Paid ads audit
- Catch wasted spend & broad-match leaks
- Find account structure gaps
- Rank your quickest wins
- Spot PMax & brand-search overlap
- Check conversion-tracking health
- Benchmark CPC vs your industry
- Catch wasted spend & broad-match leaks
- Find account structure gaps
- Rank your quickest wins
- Spot PMax & brand-search overlap
- Check conversion-tracking health
- Benchmark CPC vs your industry
Free · no credit card · instant
SEO audit
- Find keyword & ranking gaps
- Catch technical SEO issues
- Rank your fastest wins
- Surface thin & duplicate pages
- Check indexing & crawl coverage
- Compare backlinks vs competitors
- Find keyword & ranking gaps
- Catch technical SEO issues
- Rank your fastest wins
- Surface thin & duplicate pages
- Check indexing & crawl coverage
- Compare backlinks vs competitors
Free · no credit card · instant
Sarah K.
Paid Media Manager
E-commerce Agency
We started with a couple of hand-built loops for daily monitoring, then moved to Ryze for the rest. Same idea, none of the babysitting — our ROAS went from 2.4x to 4.1x in six weeks.”
4.1x
ROAS achieved
6 weeks
Time to result
95%
Less manual work
1,000+ marketers use Ryze
Automating hundreds of agencies
★★★★★4.9/5
Frequently asked questions
What is a loop in Claude Code?
A goal plus a stop condition that the agent runs on its own instead of you prompting each turn. For ads, you state an outcome once — like “every morning, flag ad sets over target CPA and draft fixes” — and Claude Code runs the discover, plan, execute, verify, repeat cycle on a schedule. The command is /loop.
Can Claude Code change my ads on its own?
Only if you connect it to the Google Ads and Meta APIs through MCP, and only within the permissions you grant. By default a loop reads data and drafts changes; keep budget and bid edits behind an approval step until you trust the gate. Ryze AI is the managed version that applies approved changes 24/7 with rollback.
When is a loop worth building?
When four things are true: the task repeats at least weekly, an objective check can reject bad output (ROAS floor, CPA ceiling, spend cap), the agent can reach the data and controls via API, and your token budget can absorb the re-runs. Miss one and a single prompt is faster and cheaper.
Which loop should I build first?
Start with a read-only one: the daily spend and anomaly watch or the weekly cross-channel digest. They deliver value immediately, carry no risk of moving spend, and teach you how the cadence feels before you let a loop change budgets or bids.
How much do loops cost to run?
Loops run on tokens, and cost compounds because context is re-read each pass — a single medium task can run 50,000 to 200,000 tokens, and a maker-checker split doubles it. Cap each loop with a token and iteration budget, and track cost per accepted change, not tokens spent.
What is the biggest risk?
A loop that moves spend you did not review. Loops fail quietly — they keep running rather than crashing — so a flawed read can shift budget unnoticed. Defend with an objective gate that can fail the work, a hard spend cap, and a human approval step before any change goes live.
Do I need to be a developer?
To build loops yourself, broadly yes — you set up MCP connectors, skills and gates, and maintain the schedule. If you want the outcome without the engineering, a managed platform like Ryze AI runs the same loops across Google, Meta and more with guardrails included.
How is a loop different from prompting for ad copy?
A prompt is one round trip you re-run yourself. A loop runs the full cycle on a schedule, checks its output against a gate, keeps a state file so it resumes instead of restarting, and stops only when a condition is met. Prompting suits one-offs; loops suit recurring, checkable work.
