Your ads, running on autopilot. You stay in charge.
Claude does the daily ad work: reads the numbers, finds ideas, makes the creatives, checks them and gets them ready. You spend 30 to 90 minutes a day approving and giving feedback. Every week it gets better from your feedback and from real booked jobs.
What your own numbers tell us
From your Meta reports, September 2024 to October 2026. When we find a winner, it carries the account. The job is finding the next one faster and cutting the losers sooner.
About C$19.06 per lead over two years. A strong base to build on.
"$999 Full House Deal" and "Pressure Washing Full House". Half of all leads.
Each spent over C$50 and brought nothing. This is the burn we want to stop sooner.
July to August 2026. More money didn't mean more leads. Decisions need daily eyes.
Three phases
Start small, prove it, then add layers. Nothing gets bigger until the step before it works.
- Claude reads Meta + GHL every morning
- Makes creatives with Higgsfield
- Your daily review page with approve / reject / comment
- Your feedback goes straight back to Claude
- Ads go into Meta as paused. You switch them on
- Every lead matched to the ad that brought it
- Booked jobs sent back to Meta so it finds more bookers
- Approved ads launch on their own, inside your budget
- Claude updates its own playbook and skills
- Weekly creative batches of 30+
- Run it every day
- Measure cost per booked job, burn, time saved
- Compare against the same months last year
- Add 1–2 add-ons that pay for themselves
- Decide: keep, grow or change
How sure are we?
Every part of this plan has a score out of 10. You'll see it on every stage, option, skill and add-on.
Build: how sure we are it can be built the way it's described, with tools that exist today.
Works: how sure we are it actually helps. More booked jobs, less wasted spend, less of your time.
Scores come from what's proven by other people, what's still new in 2026, and what depends on your account and content.
min/day
Your part
Open one page. See what happened yesterday, what Claude wants to change, and the new ads. Approve, reject, or leave a note. Hit submit. Claude does the rework. Nothing spends money without your OK, and Meta itself has a hard spending limit as a backstop.
The loop map
Two loops run together. The results loop goes around the outside every day: numbers, ideas, ads, launch, leads. The feedback loop runs through the middle: your notes and real results update Claude's playbook, so tomorrow's work is better than today's.
Click any box to open what happens there, what it costs, how long it takes and how sure we are.
Every stage at a glance
A day in the system
It reads the numbers once a day on purpose. Ads swing hour to hour, so reacting every hour wastes money. Meta's own rules watch for runaway spend in between.
Automatic runs per day
Claude Pro allows about 5 automatic runs a day. A normal day needs 2 to 4, so the MVP fits on the cheapest plan.
The week
Beyond ads: before, during and after the lead
The ad system is the core. Once it runs, the same setup (Claude + GHL + Cloudflare) can automate what happens around each lead. These are options, not all at once. Pick the ones with the best return first.
Retargeting engine
Most people don't book the first time they see an ad. Retargeting shows the right follow-up to people who already watched, asked or got a quote. GHL sorts them into groups, Meta shows the ads, and Claude writes and manages them.
A lead moves stage: new, quoted, won, lost. A GHL workflow drops them into the matching Meta audience. Built in, no code.
Each group keeps itself up to date. People leave a group the moment they book, so they never see "get a quote" again.
Ads made for that group only. Proof for people still deciding, reasons to book for people with a quote, new services for customers.
Same daily review page. Nothing new to learn.
Small budget per group. When people start seeing the same ad too often, Claude swaps in a fresh one.
Which group and which message brought booked jobs goes back into the playbook.
The groups
Each group gets its own message. The further along someone is, the more specific the ad.
| Group | Who's in it | Where it comes from | What the ads say | Leaves when |
|---|---|---|---|---|
| WarmWatched, liked, visited | Watched half a video, engaged on Facebook or Instagram, visited the site | Meta builds it on its own | Proof. Before and afters, reviews, how soft washing works | They become a lead |
| AskedLead, no quote yet | Filled the form but the quote isn't done | GHL stage: new or contacted | Easy next step. "Your free quote takes 2 minutes." Fast reply, insured, local | A quote is sent |
| QuotedQuote sent, not booked | Got a price but hasn't said yes | GHL stage: quote sent | Reasons to book now. Satisfaction guarantee, reviews, season timing, only approved offers | They book |
| CustomersJob done | Past customers | GHL stage: won | What's next. Lights for cleaning customers, yearly roof and gutter care, refer a neighbour | Never, but messages rotate by season |
| Not nowLost or went quiet | Said no, or stopped replying | GHL stage: lost | Next season, a different service, no pressure | They come back as a lead |
How Claude manages it
Week one shows proof, week two shows reviews, week three shows the seasonal reason to book. People see a story, not the same ad on repeat.
Claude watches how often each person sees an ad. When it climbs, a new version is ready in your review page before the old one wears out.
Booked and won people are removed from "get a quote" ads and from cold ads. Customers only see what's relevant to them.
Same message, everywhere they look
This is the "everywhere" effect. The same group sees one joined-up message across every place Splash N Shine shows up.
How sure we are
Built-in GHL action. Proven.
Same engine as the main ads, with a message per group.
Depends on group size. Small groups tire fast.
Google's official connector only reads today. Changes need a third-party connector.
When: about 1 week to build, once the core loop is steady in month 2. Tools cost nothing extra. Ad spend comes out of the daily budget.
Know what really works
Meta is better than anyone at deciding who sees an ad, and we won't beat that. But Meta stops at the form. It never sees who answered, who booked, what the job was worth or who came back. Splash N Shine has all of that in GHL. Put the two together and you can answer questions Meta can't.
What Meta sees
- Who saw the ad and who clicked
- Who filled the form
- Cost per lead
Great for delivery. Blind after the form.
What Splash N Shine sees
- Who answered and who went quiet
- Quote sent, booked, paid
- Job value, city and service
- Repeat customers and reviews
This is the edge. We join it to every ad.
Questions it answers every week
Not leads. Booked and paid jobs, by ad.
A full-house wash and a driveway are not the same lead.
Where to push budget, and where to stop.
Spot the slide early, before cost per booked job jumps.
Patterns across every ad, not just inside one test.
So nobody judges an ad before its leads have had time to book.
Every ad gets a label
Claude tags each ad when it's made. Over time this turns into a library of what works for Splash N Shine, and the playbook learns from it.
The building blocks
Cloudflare has the data tools for this. Start with the simple one and add the bigger ones only when the volume needs them.
1. One joined table
Meta results and GHL outcomes matched by ad, every day, in Cloudflare D1. This alone answers most of the questions above.
2. Weekly lab report
Claude writes what the data says: winners, losers, patterns, and what's too early to call. It feeds the playbook.
3. Tracking on your own website
Every page view and quote-form step recorded on the Splash N Shine domain through Cloudflare, and sent to Meta server-side. It's harder for browsers and ad blockers to lose, so Meta gets better signals.
4. Data lake
Cloudflare Pipelines, R2 Data Catalog and R2 SQL keep every event forever and make it searchable. Overkill at today's volume. Worth it once Google, website events and years of history add up.
Timeline
MVP in 1 to 2 weeks. Full loop in 4 to 6 weeks. Then 3 months of running it for real before we judge it. Each step has a test. We only move on when the test passes.
How we judge it after 3 months
Same seasons compared, so lights in fall isn't compared with cleaning in summer.
What it costs
Start at the lowest level that works, then upgrade only when you hit a limit. Ad spend is separate and stays your call. Prices are list prices in USD converted at about 1.40, so check them at signup.
Claude Pro + Higgsfield Starter + Cloudflare free. Enough to run the MVP with a few creatives a week.
When you want more daily runs and 30+ creatives a week, with more video.
Heavy video, competitor tracking, and AI texting on every new lead.
Skills and tools
A skill is a written playbook Claude follows for one job, with examples. Some already exist and are used by others. The Splash N Shine ones don't exist yet. We write them, test them on your real past data, and they keep improving from your feedback.
Future add-ons
Each one is rated on how hard it is, how long it takes to build and test, what it costs to run, and whether it actually brings more customers. Some sound exciting but won't move the needle for a local home service business. Those are marked.
What it takes to make this work
Some of this is already in place. The rest is a one-time setup. The playbook matters most, because Claude can only be as good as the rules and offers it follows.
Proof and resources
The connectors that make this possible went public in 2026. People are already running parts of this loop: Claude + Higgsfield for creatives, Claude + Meta for launching and reporting. We haven't seen anyone run the full loop for a local home service business yet. So we build it in small steps and measure everything.
What others learned, and how it shaped this plan
From guides, write-ups and videos by people already connecting Claude to Meta, Higgsfield and GHL.