A couple of years back, I spent an afternoon in Zapier trying to build an AI agent. What I actually built was an automation, and then a more complicated automation, and then a very complicated automation that I eventually walked away from. It did exactly what I told it to do every single time. That was the problem.
I think about that afternoon a lot now, because I run five podcasts and somewhere north of thirty-six episodes a quarter. At that volume, the thing that breaks is never the drafting. Drafting got solved. What breaks is the last ten feet: the post that goes out with the wrong image attached, the asset briefed as video that ships as a still, the guest promo that runs twice because two people scheduled it. Nobody catches those in the draft. They get caught by a follower, in public, on a Tuesday.
So when Loren Bartley walked me through her Claude setup, the part that stopped me was not the 150 named skills. It was the one she built specifically to tell her she was wrong.
Everything we build is designed to agree with us
Think about the assistants you have configured. Every one of them is built to produce. You ask, it delivers. You ask again, it delivers again. The entire interaction is shaped around compliance, and we have gotten so used to that shape that we stopped noticing it.
Which means the approval step is still entirely yours. You are the only thing standing between a draft and a live post, and you are reviewing your own work, on a Friday, on your ninth item. Loren said something during our conversation that I have not been able to put down. She built her checking skill the way she did, she said, because she makes a lot of mistakes.
That is not humility. That is a design requirement. If you accept that you will make mistakes at volume, and you will, then the only sane response is to put something in the workflow that is looking for them.
Where the gate actually sits
This is the part people get wrong when they try to copy it, so let me be precise about the placement.
The gate does not sit between the AI and the draft. That is quality control on the writing, and your drafting skill should already handle it. The gate sits between approved and scheduled. It runs after a human has looked at everything and said yes.
Loren’s version is a skill named Agatha, and she is the last thing that happens before content reaches Agorapulse. Loren moves rows in her Notion hub to approved, goes back to the chat she was already in, and calls Agatha. Agatha reads the approved rows and, critically, reports before she creates anything.
In our session, Agatha came back with a count first: four approved rows in the hub, three of them stageable, one that should not be touched. Then three specific things to look at.
The first one landed. The row’s visual brief said video. The only asset attached was a still. Loren went back into the hub, found the row, confirmed it, and said on camera: she had attached an image to something that should have been a video. Then, working down the rest of the list, she pulled another post entirely, because reading it back she realised she had already scheduled it earlier in the week.
She resolved all of it in conversation, not in a settings panel, not by re-editing rows one at a time. She told Agatha to skip post one, proceed with post three, leave out the duplicate, and hit return. What reached Agorapulse landed as drafts, carrying the dates her planning skill had assigned.
What you need in place to build one
Structured fields, not prose. This is the prerequisite nobody mentions. Agatha can compare the brief to the asset because the brief lives in a field. If your content plan says “make it feel energetic and visual” in a paragraph somewhere, there is nothing to check against. Your upstream planning skill has to write format, platform, offer, intended visual, and scheduled date as discrete values. The gate is only as good as the fields it can read.
Read access to both ends. The gate has to see the hub and the scheduler. Loren connects Notion and Agorapulse to Claude over MCP, which means Claude can do roughly everything in those platforms that she can. That reach is the whole benefit, and it is also the risk, and I said so on the episode. Give it read access first and live with that for a week before you give it write access.
Report before act, in the instructions. The single line that makes this work is telling the skill to return findings before creating anything. Most people write the instruction the other way around by accident: do the thing, then tell me what you did. That version is a log. It is not a gate.
A skill that runs inside the project. Agatha needs the same context the drafting skills have. As of the recording, Claude is the only tool I am aware of where a custom instruction assistant runs inside a project or notebook. A custom GPT, a Gem, and an agent in Copilot all work the same way otherwise, but none of them operate from inside the container holding your knowledge files. If you are on a different stack today, you can still build the gate. You will just be feeding it context by hand.
Be honest about what it will not catch
A gate like this is a consistency check. It compares what the row claims against what the row contains. It will catch a format that does not match its asset, a field left empty, a date in the past, a post whose platform and aspect ratio disagree.
It will not catch a claim that is false. It will not know your product changed last month. It will not tell you the post is boring. And it cannot catch a duplicate unless the scheduler is visible to it, which is exactly why Loren spotted that one herself rather than Agatha flagging it.
Keep your expectations there, and it earns its place immediately. Expect it to be a judgment layer and you will be disappointed; then you will stop reading its output, and a gate you stop reading is worse than no gate at all.
The principle underneath
We have spent three years asking what AI can produce for us. That question is close to answered. The interesting question now is what we are willing to let it withhold.
Every real team has someone who says no. The editor who sends the draft back. The ops lead who blocks the deploy. The colleague who asks whether you meant to send that. We understand instinctively that a team without one of those people ships worse work, and then we build digital crews made entirely of people who say yes.
Loren walked into that session to show us how she builds content at volume. What she actually demonstrated is that volume is safe only when something in the chain is allowed to stop it. She lost one post out of four in that batch. She was delighted about it, and she should have been, because the alternative was a video brief published as a photo to three platforms.
When you design your next AI teammate, decide early what it is allowed to refuse. That decision will do more for your output than the next model release will. The full walkthrough of Loren’s stack, all five skills, and the Notion hub they write into is in the Episode 6 breakdown, and the Blueprint there gives you the roster structure to build your own.
Pick the gate role first. Everything else on the roster is easier once something is watching the door.
