Let Claude Code run Meta ads tests end to end

One writer let Claude Code handle ad creation, campaign changes, and reporting for a newsletter experiment across 31 days.

Done with Claude Code· Advanced· Oct 2026

What they did
The writer handed Claude Code a $1,500 budget and access to the newsletter’s Meta Ads account. Once a day, they ran a single command, and the agent checked its own logs, pulled numbers from Meta and PostHog, chose whether to start, stop, or expand campaigns, carried out the changes, and saved its reasoning in git. The writer only stepped in once, to add email validation to the lead form.
How it went
During the month, it explored about 50 versions, used $1,493, and produced 243 leads. That worked out to $6.14 per lead, short of the $2.50 target. The best ads used a whiteboard-style design, and the one manual change pushed cost per lead past $50; performance never fully bounced back.
Tips
If you test this approach, make the objective open-ended instead of time-boxed, or the agent may become overly cautious. Spell out lead quality from the start so it does not wait until late in the run to account for it.

Try it yourself with Claude Code

Manage a Meta ads test for [newsletter or product] with as little hands-on work from me as possible. Draft [number] ad image and copy variations for [audience], prepare matching landing page versions, and create campaign drafts in Meta Ads through the API. Before anything is turned on, money is spent, or a form is changed in a way that could affect real users, pause and ask me to approve it. Each day, review prior run logs plus performance data from Meta and PostHog, then decide whether to launch, pause, or increase spend, and write a brief explanation of the choice in the log. Report the results by [primary metric].

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Read the original on technically.dev ↗

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.