Three things I got wrong about the Meta ads connector
Three mistakes in a week, all with the same cause: I inherited a claim instead of checking it.
I got three things wrong about this connector in a week, and all three had the same cause: I inherited a claim instead of checking it. I believed it could not see ad creatives, so I built a workaround. I repeated a tool count that was four months stale. And I described a safety guarantee as a prompting convention when it is enforced by the tooling. Each one took under a minute to check properly. The tell was identical every time — a confident, undated number or capability claim, repeated in near-identical wording across a dozen pages.
This is a young enough tool that most of what is written about it was written once, early, and then copied. That makes a specific kind of mistake very cheap to make and quite expensive to keep.
One: I built a workaround for a problem that did not exist
Every guide I read said the connector exposes text fields only — headlines, primary text, descriptions — and not the creative itself. So I did the sensible thing and worked around it: a naming convention encoding format, angle and hook, so that performance could be read without ever seeing the image.
The naming convention is fine. The reason was wrong. The connector exposes the creative directly, and the tools that do it are named here. I had spent effort solving something the platform had already solved, and I would have kept doing it indefinitely because the claim was everywhere and nobody was arguing with it.
Nobody publishes a correction when a limitation gets fixed. The original article stays up, keeps ranking, keeps getting cited, and the workaround it justified quietly becomes folklore. In a category this new, assume roughly anything described as a limitation has a shelf life.
Two: I repeated a number because everyone else did
I wrote that the connector exposed a particular number of tools. I had not counted. I had read it — in a lot of places, phrased almost identically, which felt like corroboration and was actually just copying.
When I finally counted, the real figure was several times larger, and the largest single group was catalog and product-feed tooling that essentially nobody writes about. The enumerated list is here, with the date I counted on it, because a count without a date is worth nothing in a category that ships monthly.
Three: I hedged where I should have checked
This is the one I find least comfortable, because the instinct that produced it looks like caution.
I wrote that everything created through the connector arrives paused, and I qualified it heavily — a property of how you prompt, not a guarantee, be careful, verify. That hedge felt responsible. It was wrong in the direction that costs the reader something: the creation tools declare that they create in a paused state, and none of them exposes a status parameter. It is a structural guarantee, not a habit. By hedging, I made a real safety property sound like a soft one.
Under-claiming is not automatically the safe error. It is just a different error.
The verification habits that catch this kind of thing are the least glamorous part of the routine and the part that has saved me most. See where they sit in the week →
The tell, and the check
All three had the same signature, and once you have seen it you cannot unsee it:
- A capability claim with no date on it. "It cannot do X" is a statement about a moment. Written without a date, it is a statement about a moment you cannot identify.
- Near-identical phrasing across many sources. This reads as consensus and is usually one source with copies. If several pages contain the same unusual clause, you have found one article, not several.
- A number nobody attributes. Widely repeated and never sourced means everyone got it from the same place, and nobody has been back.
The check that would have caught all three, in under a minute each:
List every tool you have available from the Meta ads connector.
Group them by what they do, give me the total, and for the
creation tools tell me what status they say they create in.That is it. One prompt, against your own account, beats any article — including this one. The reason I keep telling people to run these checks themselves is not modesty. It is that I got three things wrong in a week by not running them.
Try this tonight Run the prompt above and compare what comes back to whatever you last read about this connector. If they disagree, your account is right and the article is old. If something you have been working around turns out to be solved, that is an evening a week you get back.
What I am doing differently
Anything on this site that is a capability claim now carries the date it was checked, and where it can be checked in a single prompt, that prompt is on the page. Not because dates are a nice touch, but because I have now been wrong three times in one direction and every instance would have been caught by one minute of checking.
If you want the version of this that is a routine rather than a lesson, the skill is free and the verification habits are the first section. And if you have not connected an account yet, the setup guide now carries the corrected version of all three of these.
Frequently asked questions
Should I distrust the guides entirely?
No — most are broadly right about setup, and being early is not the same as being careless. Distrust undated capability claims specifically, because those are the ones that rot while the page stays up.
How often should I re-check?
When something surprises you, and when you are about to build a workaround. The second one is the expensive case: a workaround commits you to a limitation for as long as you keep using it.
Does this apply to anything other than tool counts?
Permissions, rate limits and rollout behaviour are all undocumented for this connector, so everything written about them — mine included — is observation rather than policy. Treat confident answers there with the most suspicion.
Why publish your own mistakes?
Partly because a page claiming to write down what actually happens should include the parts that went badly. Mostly because the pattern is more useful than the three facts: the workaround outliving the problem is going to keep happening in this category.
The habits that catch this
The Claude Ads Operator is the weekly routine these checks live inside — verify the window, ask for the tool calls, spot-check two figures before acting. Everything it drafts starts paused and budgets stay human.
See what's inside — $67 founding pricePre-sale: the modules are still being recorded and founding members get the checkout link first. No charge today. Public price $97.