Five steps, about fifteen minutes the first time. The only thing you need is export access to the ad account.
Run these in order, in one chat, with the CSV attached to the first message. Swap the bracketed bits for your account.
Attached is a [30-day] export from my [Meta] ad account for [brand]. Read it properly and give me a full breakdown of what is actually happening in this account, written for someone who knows marketing but has not looked at these numbers yet.
Cover: total spend and results with cost per result, then the same split by campaign and by ad set, so I can see where the money concentrated. Call out frequency anywhere it is climbing, CTR and CPM trends that suggest creative fatigue, and any line item where the sample is too small to draw a conclusion from. State which attribution window and date range the file is reporting on before you interpret anything, and tell me plainly which columns are missing that would have changed your read. Do not give me advice yet, I want the honest picture first.Now tell me what you would change first in this account, and why, with the specific numbers that justify it.
Rank every ad set into scale, hold or cut, and for each one name the single metric that put it in that group and the dollars involved over the period. Then give me the one change you would make first if you could only make one, the reasoning behind it, and what would have to be true for that to be the wrong call. Separately, list anything you are recommending on thin data so I can discount it, and anything that looks like a problem but is really just normal variance. I would rather have three calls you can defend than fifteen you cannot.Write me the test plan for the change we just agreed on, as something I can execute in Ads Manager this week.
Give me exactly three variants. Each one changes one variable only, and you name that variable explicitly (creative, hook, audience, placement or budget, never two at once). For each variant, tell me the budget to run it on, the minimum spend or conversion count before the result means anything, how many days it needs, and the exact kill rule: the number that, if it happens, means I turn it off without debating it. Finish with a table of what I check on day 3 and day 7, and what each outcome should make me do next.Here is this week's export for the same account, using the same date window and columns as last time.
Before you analyse it, compare it to the audit you ran last week: which of your recommendations did I actually implement, what happened to the numbers you predicted would move, and where were you wrong. Say the wrong part out loud rather than quietly rewriting the story. Then run the same breakdown and ranking as before, and finish with the one change for this week. Keep a running list of the calls you have made and how each one turned out, so we can see whether this is getting better or just getting longer.Once the weekly audit has earned its place, the next step is letting Claude read the account directly, so there is no export step and you can ask follow-up questions that need data you did not think to include in the CSV.
MCP (Model Context Protocol) is the standard that lets Claude connect to an outside service and read from it directly. Connect an ads MCP and Claude stops guessing from whatever you paste in, and starts pulling live campaign, ad set, ad and creative data from the account itself. No CSV exports, no screenshots, no "here are my numbers, what do you think".
The shift is bigger than the time saved. A dashboard tells you CPA went up. Claude tells you which three ads dragged it up, that two of them are the same creative in different placements, and that the spend moved there after your top ad hit frequency 4. Then you ask the follow-up question, which is the part a dashboard has never been able to do.
mcp.facebook.com/ads. Around 29 tools covering reporting, campaign management, catalogs and account diagnostics. No developer account, no API token, about 5 minutes.This is the one to do first even if you plan to pay for the cross-platform version later, because it costs nothing and it proves the workflow is worth it on your own numbers.
https://mcp.facebook.com/ads as the server URL. That's Meta's own hosted server, so there is no developer account, no app review and no API token to create or rotate.One connection covering five ad platforms is the actual unlock, because the questions worth asking are cross-platform ones: where does the next thousand dollars go, and which channel is quietly carrying the others.
https://mcp.gomarble.ai). Copy it from their page rather than from memory, since hosted endpoints do get moved.npx skills add gomarble-ai/gomarble-ai-paid-media-claude-skills. They give you a structured audit to start from instead of a blank chat.These only work once a connector is in place, because they ask for data no single CSV holds. They are written to be pasted as they are, with the bracketed bits swapped for your account. Each one asks for the evidence alongside the answer, because an ads recommendation you cannot trace back to a number is worth nothing.
Start here every week. This is the scale-or-cut pass:
Pull the last 30 days for [ad account name] and tell me what deserves more budget and what should be cut, using 7-day click 1-day view attribution.
Give me a table of every ad set that spent over [$X] with spend, purchases, CPA, ROAS, frequency and the trend across the last 7 days versus the 23 before it. Then split them into three groups: scale now, hold, and cut, and for each one give me the specific number that put it in that group. Flag anything where the sample is too small to call, rather than ranking it confidently on 4 conversions. Finish with the single change you would make first if you could only make one.Then the creative question, which is the one that actually compounds:
Look at my top 10 ads by ROAS over the last 90 days in [ad account name], and my bottom 10 with meaningful spend, and tell me what separates them.
For each winning ad, pull the primary text and headline and describe the hook in one line: what promise it opens with, who it speaks to, and whether it leads with pain, curiosity, proof or offer. Then find the patterns that repeat across the winners and do not appear in the losers, and rank them by how confident you are given the spend behind each one. Do not tell me "video outperforms image" unless the numbers in this account actually say that. End with three new hook angles that extend the winning pattern rather than repeating the same ad.The one nobody runs, and the one that usually finds money:
Audit [ad account name] for spend that is quietly going nowhere over the last 60 days.
Look specifically for: ad sets that spend every day and have not converted in the last 14, creatives past frequency 3 with a rising CPM and falling CTR, audiences overlapping enough that I am bidding against myself, placements taking meaningful spend at far worse CPA than the account average, and anything still live from a campaign I clearly stopped managing. For each finding, tell me the exact dollars involved over the period and what happens to my blended CPA if I remove it. Sort the whole list by dollars recoverable, biggest first.The cross-platform one, which only works if you connected more than one platform:
Compare my Meta, Google and TikTok performance over the last 30 days for [brand], then tell me where the next $1,000 of budget should go and why.
Normalise the platforms properly before you compare them: state which attribution window and conversion event each platform is reporting on, and say plainly where they are not measuring the same thing. Show spend, conversions, CPA and ROAS side by side, plus the direction each one is trending. Then give me your allocation with the reasoning, and separately tell me what would have to be true for the opposite call to be the right one. If one platform is at a ceiling rather than genuinely underperforming, say so.Competitor research, using the public Meta Ad Library:
Research what [competitor 1], [competitor 2] and [competitor 3] are currently running as ads, and compare it to what is working in my account.
For each brand, pull the ads that have been running longest, since longevity is the closest public signal we have to something working, and summarise the angles, offers, formats and hooks they lead with. Group the findings into angles they all use, angles only one of them uses, and angles nobody is using. Then tell me which of those gaps my account has evidence it could win, based on my own top performers, and which ones would be a guess. Give me three specific ad concepts from the strongest gap, written as hook plus body plus offer.And the one that saves the most time if you report to anyone:
Write my weekly paid media report for [account name], covering [date range] versus the previous equivalent period.
Structure it as: headline numbers (spend, conversions, CPA, ROAS, each with the change versus last period), what drove the change in two or three sentences of plain English, what I changed last week and what it did, what I am changing this week and why, and one risk to watch. Write for someone who does not live in Ads Manager, so no platform jargon without a short explanation. Keep it under 400 words, and put any number you are not confident in behind a clear "directional only" label instead of stating it flatly.The first answer is a report. The second, third and fourth questions are where the account actually gets better, and that's the bit a dashboard has never done. When Claude tells you an ad set is underperforming, the useful moves are: "why, specifically", "when did it start", "what else changed that week", and "what would you check before you'd bet on that explanation".
My habit is to never act on the first read. Ask what would make the recommendation wrong, ask what the smallest reversible test is, and only then touch a budget. On an account spending real money, one confident wrong answer costs more than a week of slow ones.
Meta's guide to exporting ad performance data from Ads Manager
Meta's official Ads MCP server documentation
GoMarble MCP, server URL and setup steps
GoMarble pricing and free trial
Google Ads MCP server, open source and MIT licensed
GoMarble's free Claude skills for Meta and Google audits