A marketing loop is a repeatable workflow an agent runs on a schedule, defined by nine parts: check cadence, acts-when condition, purpose, skills used, loop body, self-check, state and idempotency, stop or bail-out, and output. Match cadence to how fast the signal changes: rankings and competitor moves weekly, ad fatigue every 2–3 days, churn signals daily or on-trigger, content decay monthly. Design it so most runs check, find nothing, log "no action," and exit. Keep state in .agents/loops/ so the loop never double-acts or re-contacts someone inside a cooldown. Drafting and staging can run unattended; spending, sending, and publishing need a human checkpoint unless explicitly authorized with caps and an allowlist. Start with tracking QA and a weekly review, then add one loop at a time.
How the Marketing Loops skill works
- 1
Clarify the outcome to protect
The agent reads your product marketing context, then asks what the loop should protect or grow: rankings, ad efficiency, activation, retention, revenue, or referrals. It also checks whether this should be a loop at all. Strategy, creative direction, sparse signals like a conversion loop on 40 visitors a week, and dashboards nobody acts on are all reasons not to automate.
- 2
Pick from the 58-loop catalog
The catalog covers SEO and content, paid, earned and partnerships, outbound, activation, retention, revenue, referral and advocacy, and ongoing ops. Each entry is a complete spec, like the striking-distance push loop for pages in positions 8–20 or the failed-payment dunning loop. The agent adapts the closest match, or authors a new one from the template when nothing fits.
- 3
Tune cadence and the acts-when rule
Cadence follows signal speed, not how often you want an update, because over-frequent loops are the most common failure. The agent separates how often a loop looks from what must be true before it acts. A churn loop might check daily but act only when an account crosses a risk threshold outside its cooldown window.
- 4
Add state, guardrails, and a kill switch
Each loop gets a state file with a watermark, a handled set, cooldowns, in-flight items, and attempt counters, plus an append-only run log. Actions are split into autonomous-safe (read, analyze, draft, stage) and gated (spend, send, publish, delete). Send and spend loops get caps, suppression checks, and CAN-SPAM, GDPR, FTC, and platform-ToS rules, plus a documented kill switch.
- 5
Schedule and roll out in order
The loop body works in any agent. Scheduling uses /loop or CronCreate in Claude Code, cron elsewhere, or a manual Monday run for high-judgment loops. Rollout follows stages: tracking QA and weekly review first, then dunning and churn, then activation and conversion, then acquisition, monetization, and advocacy. Each loop has to earn its keep before the next is added.
- 6
Run a whole channel as an operator
For SEO or outbound owned end to end, the agent uses an operator recipe that bundles related loops under one schedule and shared memory in .agents/seo/ or .agents/outbound/. The SEO operator ships changes as PRs, enforces one primary query per page, and gates new pages on an information-gain sentence. Each operator ends the week with a one-screen report.
When to use it
Your agent reaches for this skill when you ask for things like:
What's inside
The playbook covers
Reference library — 7 deep-dives
See it in action
I want to set up a recurring loop that watches our SEO and tells me what to do each week. We're a B2B SaaS with a content-heavy site.
This is the SEO operator in weekly-report mode. Before the first run, settle five things in .agents/seo/brief.md: production URL, the conversion that counts (demo request or trial), 3–8 competitors, ship mode (open a PR for review by default), and data access (Search Console, analytics, rank tool). Foundation first: confirm tracking works with the tracking-QA loop and that .agents/product-marketing.md exists. Weekly loops: striking-distance push (positions 8–20 with meaningful impressions), ranking-drop watch, indexing check, internal linking, AI-answer check, and a competitor sitemap diff. Monthly: rebuild the query map, content decay, comparison-page review, and backlink prospecting.
Check cadence: weekly Acts when: a page ranks 8–20 for a query above the impression floor, or top 5 with CTR well below its neighbors Loop body: pull 28 days of query and page pairs; read the top three results and name why each outranks you; draft the smallest change that closes the gap; stage it with the starting position and CTR Self-check: is another page competing for this query? Fix the cannibalization first State: last edit date per page; wait 3–4 weeks before touching it again Stop: two pushes with no movement means escalate, since it's likely authority or intent
SEO week of [date] Shipped: each change, with its URL Moved: clicks, impressions, positions, and conversions vs. last week, each with its source Broken: open problems, worst first Next: top 3 queue items Need from you: at most 3 asks, such as data access, approvals, or facts only you have Any number that can't be sourced is left out rather than estimated.
More things to ask
Real requests the skill is tested against:
I want to automate my marketing so it basically runs itself. Where do I start?
Set up a newsjacking loop that finds trending stories and automatically posts our take to X and pitches journalists.
How do I actually schedule these loops so they run on their own?
Can you make me a loop that continuously improves our brand positioning and comes up with our big campaign ideas?
We keep losing customers to failed payments and expired cards. Set up something to recover that revenue automatically.
FAQ
How often should a marketing automation loop run?
As often as the signal actually changes. Rankings, backlinks, and competitor changes move weekly. Ad creative fatigue and CPA drift need every 2–3 days, because Meta and Google feedback loops take days. Activation is weekly so there are enough signups to be significant. Churn signals are daily or on-trigger, social mentions daily, and content decay monthly. Over-frequent loops create busywork, burn budget, and train you to ignore the output.
Can an AI agent run my marketing on autopilot?
Not without checkpoints. An agent can read data, analyze, score, draft, and stage work unattended. Spending, shifting budget, sending, publishing, and deleting records are gated behind a human by default. They run without per-action approval only when you explicitly authorize it with hard caps, a per-run change limit, and an allowlist. Crisis mentions, strategic accounts, and revenue or spend anomalies always escalate to a person.
Which marketing loops should I set up first?
Start with tracking QA and a weekly marketing review, because every other loop reads from analytics and acts on bad data if tracking is broken. Next, plug the leaks with failed-payment dunning, churn signals, and lifecycle email refresh. Then fix conversion with onboarding drop-off, signup-funnel leak, and trial conversion before turning on acquisition loops. Add one at a time and prove each earns its keep.
How do I stop an automated marketing loop from emailing the same person twice?
Give it durable state. Store each loop's memory in .agents/loops/<loop-name>.json: a watermark so it only processes new items, a handled set of dedupe keys, a cooldown map per contact, in-flight items, and attempt counters. On first run, set the watermark to now instead of processing all history. When resetting, clear the cursor but keep cooldowns, so recently contacted people aren't contacted again.
How do I know if a marketing loop is worth keeping?
Check its run log, which records items checked, items acted on, and a note for every run. If every run shows zero actions for weeks and nobody misses it, it's a vanity loop and should be deleted. If it acts on every run, it's probably chasing noise. Re-audit quarterly: recalibrate thresholds, retire loops nobody acts on, and keep only as many loops as you can review.
What should every recurring marketing workflow include?
All nine parts: check cadence, acts-when condition, purpose, skills used, loop body, self-check, state and idempotency, stop or bail-out, and output. The self-check, state, and stop condition matter most. The self-check rules out noise, seasonality, or a tracking bug before acting. State prevents double-acting. The stop condition caps attempts, handles errors, and hands off to a human. If you can't answer those three concretely, the loop isn't ready.
Related skills
Use it in your agent
Marketing Loops is a plain SKILL.md folder, so the same skill loads in any agent that reads the skills format. Setup guides:
Build a tool for this workflow?
Marketers run this skill mid-task — the exact moment your tool matters. Partner tiers put your tool inside the skill their agent runs.
See partner tiers