The Review Gate: How We Quality Control 88 AI Fellow Videos in Two Weeks | Sanjana

Over the last two weeks of September, 88 fellow videos landed on the desk of Sanjana, a project manager at Humanitarians AI. The video is her record of what she actually did, how she checked each one, and what made it through. Her central point is that reviewing is not watching a video and clicking approve. Every submission runs a gate first.

The three checks every video must clear

The gate has three checks, and every video faced the same ones before reaching a professor. The first asks whether the video is rendered in true 4K, both the wide cut and the vertical short, not just one of them. The second asks whether the branding is right: the Humanitarians AI handle should appear, not a leftover personal handle from an earlier template. The third asks whether the code is on GitHub with the dimensions correct. Three passes and the video is approved and goes to the professor. Any fail and it goes back to the fellow with an email saying exactly what to fix.

Turning two weeks of notes into a scoreboard

Two weeks of review notes make a spreadsheet, not a picture. So Sanjana did what she says she would tell any fellow to do: she handed the raw tracker to Claude and asked it to turn the mess into a scoreboard. The request was not to judge anything, only to count and lay the numbers out honestly. The scoreboard showed 88 videos reviewed from 22 different fellows. Of those, 78 cleared the gate and went up for approval, 10 came back needing fixes, and 71 already had their code on GitHub where the work can be checked.

Where the videos came from

The 88 were not spread evenly. Almost half, 41, were Brutalist Toolkit videos as that project scaled up. Madison came next with 18, then the AI plus one teams with 14. The remainder were spread thinly across several smaller projects, eight projects in all for one reviewer.

What a clean pass and a returned batch look like

The video walks through the gate working on a good submission. One fellow's batch of four brutalist videos passed on all three checks, with 4K rendering, correct handle branding and code on GitHub, so it was approved and uploaded to YouTube for the professors with no email and no back and forth. Most of the work, however, is in the ones that fail. In another example, a batch of 16 videos had good content but needed changes, so it went back with notes. The next day the fellow fixed them and re-uploaded, and the batch was approved. Other returned batches followed the same pattern. Sanjana's point is that a gate is not meant to reject people. It exists to catch problems, say exactly what is wrong, and let fellows fix them before a professor ever sees the work.

Build your own gate

The closing advice applies to anyone who reviews work, their own included: build a gate instead of making a gut call. Paste your log into Claude and have it turn the log into a scoreboard and a checklist you apply to every item, the same way every time. The checklist is the real output. A good answer gives you three or four checks you can actually run, while a vague answer that just says "looks good" is the gut call you are trying to replace.

Key takeaways

• Reviewing a submission means running it through defined checks, not watching and approving.
• The gate used here has three checks: true 4K rendering of both the wide and vertical cuts, correct Humanitarians AI branding, and code on GitHub with correct dimensions.
• A video that passes all three goes to the professor; any fail returns to the fellow with specific fixes.
• Claude can turn a raw tracker into an honest scoreboard when asked to count rather than judge.
• In the two-week window, 78 of 88 videos cleared the gate and 10 came back for fixes.
• The purpose of a gate is to catch problems early and give clear feedback, not to reject people.
• A useful checklist has three or four runnable checks, not a vague verdict.

Try it yourself

If you review anyone's work, paste your review log into Claude and ask it to produce a scoreboard and a checklist you can apply to every item in the same way. Then read what it hands back and check that the checklist has three or four checks you could actually run, rather than a general impression.