How a nightly AI series runs on TikTok without a camera

The Vibe Coders PH TikTok has 2,846 followers and 52,100 likes at the time of writing, and I have never filmed a frame for it. Every post is written, voiced, rendered and scheduled by a pipeline, and the part I actually control is what gets made and what gets killed.

This is the build log for that pipeline: the two daily slots, what each format is, how a piece travels from an idea to a scheduled post, and the single number I use to decide whether a format earns its place.

Two slots, two jobs

The channel runs two posts a day, and they are not the same kind of thing.

19:50, Claude Kami. A Taglish skit where I argue with Claude at a sari-sari store. This is the expensive slot: a character, a set, a voice, a joke that has to land in three seconds. It is the reason people follow.

12:30, the cheap formats. Araw N is a build day in one short. Sulit ba? is a tool judged on whether it is worth the money. On Sundays a carousel, Capstone na Papasa, answers the questions students actually ask before a defense.

The split is deliberate. The evening slot is the one that builds a character, so it cannot run every day at a quality worth watching. The midday slot is cheap enough to run daily, which makes it the place where a format gets tested. If a cheap format earns follows, it graduates. If it does not, it dies without touching the character.

What each format actually is

Claude Kami is code. The set, the character, the sari-sari store shelves, the text on the wall are all drawn in a 3D scene by a script, then recorded frame by frame. Nothing is shot. The first episode went out on 6 October 2026 with the line "Tapos na yung app. Sino nang gumagamit?" and the joke is the one every builder knows: you shipped, and nobody came.

PATAY GUTOM is a serial of security-camera horror films at real Philippine places. The EDSA busway at 2 AM, a last MRT-3 train, the NAIA terminal, a Megamall after closing, the Araneta floor on a finals night. Every frame is a pure function of time in a Three.js scene, which means a film renders identically on any machine. That property matters more than it sounds: when a shot is wrong I fix the code and re-render, instead of hunting for a take that will never come back.

Explainer Lab is paper-cutout animation, drawn frame by frame in code, answering Filipino questions: why so many nurses are Filipino, when Filipinos first reached America, what a mameng is. It publishes across Facebook, Instagram and TikTok.

How one piece gets out

1. A search term comes first, not an idea. Araw N exists because people search "paano gumawa ng website gamit AI". Sulit ba? exists because people search "sulit ba ang Claude Pro". The first line of the video and the first 50 characters of the caption both carry that term, because a recommendation only helps if the person already wants the answer.
2. Script. Written with Claude, reviewed by me. The rule is that the opening line states the viewer's stake, not my credentials.
3. Voice. ElevenLabs, cloned. The voice is consistent across every format, which is most of what makes a channel feel like a channel.
4. Render. Code. For films, the output is a frame sequence plus a cut list: which shots, in what order, with which overlays and signs. For skits, a scene file and a render pass.
5. Cut and caption. ffmpeg does the assembly. The caption carries the search term, the disclosure if the piece is AI-generated, and the next episode's hook.
6. Schedule. Buffer holds the slot. The autopilot keeps at most seven episodes ready, and nothing posts without my yes.

That last point is the one I will not automate. The pipeline can write, render, caption and queue. It cannot decide what is worth saying.

The only number that matters

Views are a vanity metric on TikTok. Follows per 1,000 views is the real one.

Episode 01 of Claude Kami was read at 15 hours with 4,810 views, 10.4 follows per 1,000 views, 64 likes per 1,000 and 15 shares per 1,000. It brought 50 new followers in 15 hours, against an account that had 175 at the time. Twenty-nine percent of the audience arrived from one video.

Every piece is read twice: at 24 hours for the hook, at 7 days for whether it keeps working. The four-week test that started on 12 October is judged on the same number. A format that wins views and loses follows is a format that gets cut.

The first week also settled an old argument. TikTok carried 81 percent of the views across the channel at 27 views per follower, against 4 on Instagram and 0.3 on Facebook. The platform that was hardest to make for was the only one doing the work.

What broke

The promo opening. The organiser-first promo and the long English explainers lose more than 40 percent of viewers in the first second and bring almost no follows. Both formats are retired. The fix was not better editing, it was opening on the viewer's stake.

The queue outran the cap. On 9 October the autopilot was eight posts ahead on TikTok with one free slot, so it stopped making new work until the backlog cleared. A cap only works if the queue respects it, and the queue is a calendar, not a count.

The expensive slot drifts. The evening piece wants to be a short film. It is a 40 second skit. Every week the scope creeps and every week the render time has to be argued back down.

What I would tell someone building one

• Pick the number before the format. I picked follows per 1,000 views, which immediately killed two formats that looked fine in a dashboard.
• Make the cheapest format the daily one and the best one the weekly one. Running your best idea every day is how you burn it.
• Make the renders deterministic. A film that cannot be reproduced is a film that cannot be fixed.
• Automate everything except the yes. The pipeline should be able to produce a week of work overnight, and you should still be the one who decides it goes out.