Run a one-person app's marketing with a coding agent and connectors

A solo iOS developer had a coding agent build a bilingual content pipeline, drive AI video tools and set up a Meta ad campaign that cut cost per thousand views from $8.20 to about $2.70.

Done with

Custom agent

Apps touched

What they did

The author makes TransApp, a one-person Apple app that transcribes WhatsApp voice notes. Between May and September 2026 they used an unnamed AI coding agent, a git repo and several connectors as the marketing team. Two repo files hold the brand voice and design rules, and every generator reads them first. The agent built Python and Pillow post renderers in English and Portuguese, an HTML carousel engine exported with Playwright, a Remotion motion project, a redaction pipeline for real chat footage, and set up profiles on nine platforms through a browser. It drove Higgsfield's web interface and MCP server to make vertical video ads, built a Meta traffic campaign end to end through Meta's Ads MCP server, filed assets in Google Drive, and pulled numbers from server logs, App Store Connect, Search Console, web analytics and Meta to compare them in one table.

How it went

Two of about fifteen first-round video clips were usable; the final shot worked only after the agent warped a real app screenshot into the clip's last frame to use as the end frame. The first Meta test bought views at about $2.70 per thousand against $8.20 for the boosted post it replaced, with roughly one in ten tapping through at about 3 cents per landing-page visit. Comparing sources showed almost nobody was reaching the App Store page. The Instagram account is still small, and the author says this is not a results post.

Worth knowing

Account creation, passwords, two-factor, captchas, spending money and posting on Reddit and X stayed manual. Meta rejects App Store links outside app-install campaigns only in delivery diagnostics minutes later, so the agent now checks those before reporting success. Some web forms drop values on submit, so every fill ends with a read-back. The agent product is not named.

Try it yourself withTry it withCustom agent

10 parts to fill in

Act as the marketing team for [my product]. First read [brand voice file] and [design file] and follow their rules in everything you make. Render this week's [number] posts in [languages] at [size], open each rendered image and check it against the rules before saving it to [shared folder]. Then draft a [traffic] campaign in [ad platform] with UTM tags pointing at [landing page], read its delivery diagnostics, and stop and show me the campaign and a proposed budget before anything is published or any money is spent.