Jacob Malherbe

Managed $ 250M+ in Meta ad spend since 2015. Now I build MCP servers and AI agents on Cloudflare (~270 in production). Shipping YAB-E. Father of 5.

[Jacob Malherbe @Jacobmalherbe21]

I've run Facebook ads since 2015.

Most builders I meet have never run one.

So I wrote down the 5 things it took me 10 years to learn. Takes 3 minutes.

What's stopping you from running your first ad? https://t.co/yHwgzwdtsD

[Jacob Malherbe @Jacobmalherbe21]

@gregtosi Ads pour gas on a fire. If there’s no fire yet, you’re just buying an expensive way to find out.

[Jacob Malherbe @Jacobmalherbe21]

I've run Facebook ads since 2015.

Most builders I meet have never run one.

So I wrote down the 5 things it took me 10 years to learn. Takes 3 minutes.

What's stopping you from running your first ad? https://t.co/yHwgzwdtsD

[Jacob Malherbe @Jacobmalherbe21]

@grwth_hackr following along

[Jacob Malherbe @Jacobmalherbe21]

@marckohlbrugge Add to the list. https://t.co/4iJeZ3ZMXE

[Jacob Malherbe @Jacobmalherbe21]

@larsbuilds I think Muse will be huge in the future so I am building out https://t.co/7BZqoduBTG

[Jacob Malherbe @Jacobmalherbe21]

@PhedEU Manually picking 8,000+ thumbnails is the quiet moat here. Curated beats scraped every time, and it should make that MCP far more useful when an agent calls it.

https://t.co/4iJeZ3ZMXE

[Jacob Malherbe @Jacobmalherbe21]

@matchaman11 Works with existing subscriptions is the detail that sells it. Nobody wants another bill, and that removes the friction before anyone even tries it. Nice work shipping all three platforms.

[Jacob Malherbe @Jacobmalherbe21]

@devsolojr Daily is the right rhythm. Showing up every day is what carries a builder when the outcome is unsure. One lesson from my builds: let one real founder's hard week shape the messages.

[Jacob Malherbe @Jacobmalherbe21]

@ethan_kinnan Where agents get stuck is the useful part. First fix to try: add an /mcp endpoint so they skip the crawl. Sherpa can also list free on Launch via MCP, straight from an AI chat.

[Jacob Malherbe @Jacobmalherbe21]

@Gregoirechp Ten months on scrapers is the moat, and shipping early because people asked is the signal. Stay on it. You can also list it free on Launch via MCP, straight from your AI assistant.

https://t.co/4iJeZ3ZMXE

[Jacob Malherbe @Jacobmalherbe21]

@shubham_19hshs When Meelu checks if ChatGPT, Claude and AI Overview cite a site, how do you handle answers that change from run to run? Also, Meelu can be listed free on Launch via MCP from your AI assistant.

https://t.co/4iJeZ3ZMXE

[Jacob Malherbe @Jacobmalherbe21]

@sanathbhat39 Unapproved refunds is the example that lands, since every owner feels it. Tip: turn each refund rule into a test, then try many ways of asking. You can also list it free on Launch via MCP.

https://t.co/4iJeZ40kNc

[Jacob Malherbe @Jacobmalherbe21]

@okshubhh Reporting only changes since the last check is the detail that matters. Same lesson from my Workers: a registry Claude checks before any edit stops the fixed-one-thing-broke-three problem.

https://t.co/4iJeZ3ZMXE

[Jacob Malherbe @Jacobmalherbe21]

@marclou Putting alerts behind the paywall is the smart split. A globe gets admired once, but an alert pulls people back every day, and that habit is what turns free users into paying ones.

[Jacob Malherbe @Jacobmalherbe21]

@DanielSmidstrup All in with no guarantees takes nerve. Faith keeps me in the room when builds break. Keep going, and let customer one be your roadmap.

[Jacob Malherbe @Jacobmalherbe21]

@beamnxw Durable-harness is the sleeper on this list. Resuming agents after an interruption beats starting from zero every time. Thanks for doing the scan.

[Jacob Malherbe @Jacobmalherbe21]

@marclou The learn half paid me most. Google's Panda update killed my SEO approach, and that loss pushed me into Facebook ads. Keep going, every lesson compounds.

[Jacob Malherbe @Jacobmalherbe21]

@jdudley @AnthropicAI @xai @Meta @OpenAI Interesting that Muse ranks above Duet for you. Is the gap in the connectors it can reach, or in how it runs on its own once you leave it alone?

[Jacob Malherbe @Jacobmalherbe21]

@Thisaitooldoest Capping it at 3 is the quiet win. A limit forces Muse to rank instead of dump, so the Notion database stays something you actually read at 7:05 instead of a pile you ignore.

[Jacob Malherbe @Jacobmalherbe21]

@uttamvarma The cohesive loop has to include the connectors too. One dead or slow Muse connector and the whole harness feels broken, however good the model is. Health checks keep that frictionless feel.