A $2,019 Claude Code Meta Ads Test: The 7.5x Arithmetic Holds, the Repo Contradicts the Hands-Off Claim

The video claims a content creator handed Claude Code her existing video library, a Meta API key, and her credit card, and that over seven days the agent spent $2,019.18 and produced $15,235 in gross sales. The division is correct, and the word attached to it is wrong: 7.5x is a platform-attributed revenue multiple on media spend, and the repository the video recommends states that every ad it creates launches paused for a human to review. On screen there is a red banner reading "CREATOR SOLD $15K USING CLAUDE X META," a green overlay reading "+$15,235.00 (7.5x return)," and a picture-in-picture of the narrator reacting over a clip of a woman at a laptop. Later frames show a Claude Code terminal with a repo URL and the words "help me set this up," an Ads Manager table filmed off a monitor, and a GitHub file listing. One frame carries the on-screen question "did you give it a limit on what it could spend?" The video never answers it.

The multiple checks out

15,235 divided by 2,019.18 is 7.545. Rounded to one decimal that is 7.5. The headline multiple is arithmetically correct, which is worth establishing before the rest of this takes the claim apart, because plenty of viral ad-results videos fail at that first step. The subtraction is equally simple: $15,235 minus $2,019.18 leaves $13,215.82 of gross sales above media cost, a 6.5x net multiple on the budget.

"Return" is the word doing the damage. A 7.5x return implies you keep 7.5 times what you put in. What Ads Manager reports is return on ad spend, which counts exactly one input. The product here is physical hardware. Cost of goods, shipping, payment processing, and returns all come out of that $15,235 before anyone reaches profit. The video shows the creator's own prompt on screen, and that prompt includes the line "if you need any additional context about COGS or anything like that, let me know." Cost of goods was on her mind while briefing the agent. It appears nowhere in the result the video reports.

Gross sales on a screenshot is platform-attributed, not audited

Meta Ads Manager reports the conversions its own attribution model assigns to your ads. The default window is 7-day click and 1-day view. Under the view half of that, a purchase counts toward Meta when the buyer saw the ad without clicking it and then bought through any other path within 24 hours. Geo-lift testing that compares platform-reported ROAS against measured incremental ROAS commonly finds gaps in the range of 1.5x to 3x, widest on retargeting and on audiences that were going to convert anyway.

That caveat carries unusual weight here. The clip shows an Instagram post with 183K likes, and the creator's prompt says she has "always posted organic content" and holds "a library of high-performing organic content." She was already going viral when the test started. Organic demand that converts during a paid window lands in the paid column. Nobody in the video mentions a holdout group, a geo test, or a comparison against the prior week's baseline sales, so there is no way to separate what the ads caused from what was already happening.

The repo's own documentation contradicts the hands-off claim

The repository is krusemediallc/arcads-claude-code. A terminal frame shows a URL ending in "cads-claude-code," and a later frame shows the GitHub owner krusemediallc above a file listing of .claude, skills, scripts, AGENTS.md, and an MIT license. Its meta-ad-builder skill documents its own behavior directly: "Every ad is created PAUSED. Tell the user to review and un-pause in Meta Ads Manager." The same file states that the skill creates ads only, and that it does not create or edit campaigns, ad sets, budgets, or audiences.

Read those two lines against the video's claim that the setup "made it so I didn't have to click a single button in Ads Manager." Somebody has to build the campaign, build the ad set, set the budget, choose the audience, then open Ads Manager and unpause each ad. That is a person doing Ads Manager work both before and after the agent runs. The agent writes copy and assembles creative into ad objects.

The credit card framing is off in the same direction. The repo asks for two environment variables, META_ACCESS_TOKEN carrying the ads_management scope and META_AD_ACCOUNT_ID. The card sits with Meta as the ad account's payment method, and the agent never sees it. What the agent receives is authority to create objects inside an account that bills a card already on file. The exposure is real, and the mechanism differs from the one described, which matters when you try to reason about what could go wrong.

The repo belongs to the narrator, and it has more stars than he said

@mr.paidsocial is Caleb Kruse of Kruse Media LLC. The GitHub account krusemediallc is the same Caleb Kruse. The Ads Manager account visible during the Skool segment is labeled "Mr. Paid Social." The video says "go to this github repo" without noting that the repo is his own, and he sells a paid Skool community built around these workflows. None of that makes the code bad. It does turn "the best way I've seen to run them right now" into a recommendation with an interest behind it.

The star count is understated. The video says 1,100. As of 2026-09-29, two days after the video posted, the GitHub API reports 1,552 stars and 376 forks on a repo created 2026-04-01. Star counts move daily and this one is climbing, so read any specific figure as a timestamp. Stars measure bookmarking, and they carry no information about whether the code produces profitable campaigns.

One cost the video omits entirely: the repo describes itself as an "Arcads external API" workspace and requires an Arcads API key, with Meta credentials listed as optional and needed only for publishing. Arcads is a paid subscription product with plans running from roughly $29 to several hundred dollars a month. Neither that subscription nor a Claude Code subscription sits inside the $2,019.18.

Seven days with one creator is a sample of one

$2,019.18 across seven days averages about $288 a day. Meta's learning phase targets roughly 50 optimization events per week per ad set before delivery stabilizes, and performance inside that window swings hard by design while the system explores placements and audiences. A seven-day test is one learning phase. It is the noisiest week the campaign will ever have, and the reported figure is wherever that noise happened to settle.

There is no control, no repeat run, and no losing run anywhere in the video. Accounts produce both. The narrator's own Skool-sale screenshot shows the shape of it: five ad sets, two of which spent money and returned a conversion value of $0.00. The rest of that table is filmed off a monitor at an angle and does not resolve on playback, so the $30,000 figure attached to it cannot be verified from the video at all. The community platform is Skool, spelled with a K, which the auto-generated transcript renders as "school."

The narrator's credentials follow the same pattern. Twelve years running Meta ads and over $150 million in platform spend are self-reported, repeated across his own LinkedIn and course marketing, and unauditable. Meta publishes no per-manager lifetime spend figure that anyone could check them against.

The failure modes worth planning for

A frame in this video displays the question "did you give it a limit on what it could spend?" and then moves past it. Answer that before running any of this. Daily and lifetime budget caps live at the campaign and ad set level, which this repo explicitly does not touch, so those caps are yours to set. The harder stop is an account spend limit on the ad account itself, and that is a manual setting a human configures once.

Meta's Platform Terms require you to protect access tokens and state that you must not transfer or share them. Running an agent on your own machine against a token you generated for your own ad account is you using your own credentials, which is what the terms contemplate. Pasting that same token into a hosted service you do not control is the version that creates a problem. Separately, Meta supports Marketing API automation on connections you authorized yourself, while automations that generate large volumes of machine-created activity in a short span, or that drive a logged-in Ads Manager session through a browser, are the pattern associated with accounts disabled for unusual activity. Creating a batch of paused ads through the API sits on the supported side of that line. Spinning up and launching dozens of campaigns an hour on a young account does not.

Key Takeaways

15,235 divided by 2,019.18 equals 7.545, so the video's 7.5x figure is arithmetically correct. The label "return" is what fails, because ROAS on gross sales is not profit.
Subtract cost of goods, shipping, and payment processing before treating any of the $13,215.82 gap as margin. The creator's own on-screen prompt raises COGS and the reported result never accounts for it.
Meta's default 7-day click and 1-day view attribution credits Meta for purchases it may not have caused, and this creator had a large organic audience converting inside the same window.
The repo is krusemediallc/arcads-claude-code, MIT licensed, owned by the narrator, with 1,552 stars and 376 forks as of 2026-09-29 rather than the 1,100 stated.
The meta-ad-builder skill creates ads paused and does not create campaigns, ad sets, budgets, or audiences, so "didn't click a single button in Ads Manager" does not match what the tool does.
Set an ad account spend limit and ad set budget caps yourself before handing any agent an ads_management token, because this repo manages neither.
Budget for an Arcads subscription and a Claude Code subscription on top of media spend, since the repo requires an Arcads API key and the $2,019.18 excludes both.
Judge the setup on a month of runs across several products rather than one seven-day window that sat entirely inside Meta's learning phase.

Resources

krusemediallc/arcads-claude-code. The repository the video points to, owned by the narrator, MIT licensed, 1,552 stars as of 2026-09-29.
meta-ad-builder SKILL.md. The skill file stating that every ad is created paused and that the skill does not create campaigns, ad sets, budgets, or audiences.
Meta Platform Terms. Section 6 covers safeguarding Platform Data and the requirement not to transfer or share access tokens.
About the Learning Phase. Meta's own explanation of why delivery and cost stay unstable until an ad set accumulates enough optimization events.
Marketing API Access Tier update. Meta's developer blog on the access tiers that gate how much API-driven ad management an app can perform.
Arcads. The paid AI video ad product whose API key the repo requires, and whose cost sits outside the $2,019.18 figure.