Blog

A TikTok channel run 100% by AI agents (and it shows)

Agents orchestrated by SideHub, my personal tool, write, edit and publish three videos a day on @electrochoc.mind with no human in the loop. The videos are not worth much. The machine behind them is: the pipeline, the shared memory, what broke and what I learned.

Published
Line
AI
Reading time
9 min
Languages
EN · FR

Nobody at the keyboard (and you can tell)

This morning at 9 a.m. Paris time, a video went live on @electrochoc.mind. In French, it says: “You’re running from your urgent file. You keep Instagram open to avoid your urgent file. Lock Instagram then display your urgent file now!”

Read it again, slowly. Nobody runs from a file. Nobody “displays” a file the way you face your demons. And nobody has ever locked Instagram in a burst of courage. The grammar holds up. The meaning, not so much. 🙃

I didn’t write it, edit it or publish it. I didn’t even see it before it went out, which, as you may have guessed, shows. Two more will follow at 2 p.m. and 8 p.m., as they have every day since May.

Let me be upfront: this is not a coaching channel I am proud of. The videos are not good, and I know it. What interests me is everything running behind them. This channel is my playground for SideHub, the tool I built for myself to run my own coding agents. I am its only user, and it is not for sale. The question was never “can agents make good content?”. It was: “can agents sustain a daily production for months without me stepping in?”. The answer is yes. Quality is another story. 😅

Here is how it is built, what broke, and what I took away from it.

The format: three acts in twenty seconds

Each video is meant to be an “electroshock”: an uncomfortable truth, told in French, in the second person, in a harsh tone. It lasts 15 to 22 seconds, in vertical format, and always follows three acts:

  • the hook, 5 to 10 words, meant to stop the thumb
  • the tension, 8 to 16 words, naming what you would rather not see
  • the action, 5 to 14 words, a concrete order that always ends with an exclamation mark.

Each act has its own background clip, and a 60-millisecond white flash marks the cut. Six voices share the scripts (the brutal mentor, the disappointed father, your future self, the soldier…). Yes, a synthetic disappointed father orders you to close Instagram. 🫠 About fifteen visual settings rotate from one video to the next: gloomy everyday life, old age, the sea, sport…

These constraints are not about style. A strict format is what makes automated production controllable. Every rule can be checked: word count, duration, the exclamation mark. Meaning cannot. That is the whole problem, and we will come back to it.

The production line

SideHub scheduler (9 a.m., 2 p.m., 8 p.m. Paris)
  └─ workflow "Generate new video"
       1. ideas       ~10 scored candidates, one picked
       2. checks      blocking editorial gate
       3. voice       ElevenLabs (OpenAI TTS as fallback)
       4. footage     Pexels (video) + Jamendo (music)
       5. render      HyperFrames → 1080×1920 MP4
systemd timer (hourly)
  └─ publishing     Zernio → TikTok, then SQLite registry

All the know-how lives in a Claude Code skill: a SKILL.md file describing the procedure, an editorial guide and a handful of small Node.js tools. The agent reads the skill, calls the tools in order and makes the creative decisions. The tools do the mechanical work. Judging by the output, the tools are doing better.

Ideas. The first workflow step produces about ten candidates. Each one is scored on several axes: avoiding repeats, following what performs, hook strength, freshness of the action, narrative consistency. The step also lists overused verbs and neglected personas or settings, to force variety. A second step picks the winner. So this morning’s video was the best of ten. I try not to think about the other nine. 🙈

The editorial gate. An agent writing three scripts a day ends up repeating itself. Mine loved deadlines: “tonight”, “before midnight”, “before you go to sleep”. Instead of politely asking again in the prompt, I wrote a deterministic check that blocks the script:

const DEADLINE_RE = /(ce soir|avant ?ce soir|avant minuit|avant de dormir|cette nuit|demain matin|…)/i;

checks.action_exclamation = /!$/.test(action);
if (!checks.action_exclamation) reasons.push('action doit finir par !');

The same tool rejects out-of-range lengths, pronouns with no referent (“call him”, “she”) and metaphorical actions nobody can actually do. It exits with an error, and the agent has to rewrite.

Notice what is missing from that list: nothing checks that the sentence means anything. This morning’s video passed every rule with flying colours. Right word count, exclamation mark, no deadline, no orphan pronoun. It is simply absurd. 🤷 A regular expression can tell whether a sentence ends with “!”. It cannot tell whether that sentence makes sense.

Voice and footage. ElevenLabs reads the script in the persona’s voice, and OpenAI takes over if it fails. An ffmpeg pass adjusts tempo and loudness. Pexels supplies three clips, one per act, picked from queries allowed for the day’s setting. A blacklist rules out images that have become clichés, which is a bit rich given the scripts laid on top. Jamendo supplies the music.

Rendering. The compose tool measures each act in the audio, fills an HTML template and renders it with HyperFrames, which turns an animated HTML page into a video. It takes one to three minutes per video. It is by far the most reliable step in the whole chain.

Orchestration

Without orchestration, all of this is a script someone has to run by hand. In my case SideHub handles it, simply because I already use it every day for my agents. Any tool that can schedule jobs and share state would do. It takes four things.

  • Schedulers. One triggers the generation workflow three times a day (0 7,12,18 * * * UTC). Another fetches TikTok statistics every four hours, offset by thirty minutes so it never overlaps with generation.
  • The workflow. Each step is a separate agent session with its own prompt. It drops its output in the Drive and signals completion with sidehub-cli workflow step-complete. Without that signal the workflow does not move on, and the step times out after a set delay.
  • Shared memory. A SQLite database in the Drive acts as the registry. The published table keeps every published video (scripts, persona, setting, clips used), so the ideas step never writes the same thing twice. It writes something new every time, not necessarily something better. The stats table tracks views, likes, shares and engagement rate for each video. Several agents write to it without stepping on each other, because the backend serializes writes.
  • Alerts. A state page counts consecutive failures. On the second one, the agent creates a task in SideHub, and I see it on my phone.
SELECT universe, COUNT(*) AS videos, ROUND(AVG(s.views)) AS avg_views
FROM published p JOIN stats s ON s.tiktok_post_id = p.tiktok_post_id
GROUP BY universe
ORDER BY avg_views DESC;

The feedback loop

The statistics are not just there to reassure me, which is just as well, because they are not very reassuring. On every run, the agent gets a “performance brief” built from the stats table: the settings, personas and hook shapes that work best. Today, the “old age” setting averages about 470 views per video, well above the channel average, and the “future self” persona holds the channel’s two best videos. Moral of the story: being lectured by yourself forty years from now works better than being lectured by a soldier. 👴

Loops that feed on themselves still deserve suspicion. Zernio, the publishing tool, suggests the “best times” to post. But it only measures the time slots already used, so it always confirms my choices. The rule now is to test a new slot every thirty days, to give it data it doesn’t have.

What broke

On the content side, as we have seen, meaning breaks on a regular basis. But the technical side is what taught me the most.

One backend too many. The first version, in August 2025, was a C# and Blazor app: Creatomate webhooks for rendering, Hangfire for scheduling, scraping for ideas. It worked, but every format change meant code. On May 6, 2026, I deleted it in favour of a skill and a few command-line tools. Changing the format now means editing a Markdown file.

Publishing moved out of the agent. At first, the last workflow step published the video. But that step runs in an agent in autonomous mode, and an agent can refuse an action or get blocked by its own safeguards. For a task as mechanical as an upload, that risk brought nothing. A plain systemd timer now runs publish.sh every hour. The script takes the next ready video, publishes it and records it in the registry. It never wonders whether the video means anything, which is exactly what we ask of it.

Unreadable videos. Some Pexels clips are HDR. HyperFrames then switched to HEVC, and the video check, which expects H.264, failed the step. The --sdr flag now forces H.264. Unreadable in the technical sense, for once. 😏

// --sdr: force SDR/H.264, otherwise HyperFrames switches to HEVC on HDR sources
const renderArgs = ['--yes', 'hyperframes', 'render', '-o', 'out.mp4', '--sdr'];

Safeguards at publishing time. The editorial gate runs again right before publishing, and an invalid video is skipped instead of blocking the queue. A .published.json marker file, written after each post, prevents any republishing. Upload and post creation each get one retry after thirty seconds.

The lesson fits in one sentence: creative work to agents, deterministic work to plain code. Agents pick the idea, write and make the calls. Everything that must happen the same way every time (checking, rendering, publishing, recording) is a script that doesn’t think. Judging by the scripts, the creative part probably deserves a human. But that was not the experiment.

The numbers, honestly

  • Over 200 videos published since May 2026, three a day.
  • About 270 views per video on average over the last fifty, 802 for the best one.
  • 5 to 10 euro cents of voice per video. Footage and music come from free APIs.
  • One to three minutes of rendering per video.

It is not viral, and that was never the goal. Luckily. What matters is that the system has run on its own for months: it publishes, measures, adjusts and warns me when it gets stuck.

Last stop

This channel produces forgettable videos. Yet it taught me more about agent orchestration than any demo. A single agent is brilliant at a task. To sustain production over time, it needs what any team gets: a schedule, a shared memory, checks that don’t depend on its mood, and someone to warn when things break. For me, SideHub plays that role, a tailor-made tool whose only customer is me.

Producing mediocrity reliably, three times a day, for months, is still a real engineering problem. The day agents write sentences that actually mean something, the plumbing will be ready. 🔧

Next departure

If I ever wanted these videos to mean something (one can dream), here are the leads I would try.

  • A reviewer who understands French. The editorial gate checks form, never substance. What is missing is a step where a second model reads the script like a viewer and answers one question: “would a human say this?”. That check would no longer be deterministic, but it is exactly the one that is lacking.
  • Real examples instead of rules. Today the agent mostly learns what not to do. Showing it a handful of short, well-written scripts written by hand would finally give it an idea of what to do.
  • Loosening the format. Three acts, fixed word counts and an action that ends with “!”: these constraints make production controllable, but they also push the agent to fill in boxes. A looser format would give more natural scripts, at the cost of less simple checks.
  • Measuring watch time, not just views. The registry tracks views, likes and shares. Watch time, if it is available, would show when people drop off, and therefore which sentence doesn’t land.
  • A better writing model. This is the most obvious lead, and the most expensive. A high-end model would probably write more sensible sentences, but the ideas step runs three times a day on about ten candidates. The bill would climb fast for a channel that earns nothing.
  • Footage that matches the script. Pexels clips are picked per setting, not per sentence. Generating tailored shots with a video model would make the whole thing more coherent, but the cost per video would jump by an order of magnitude.

There is always the radical option: writing the scripts myself. But then it would no longer be a channel run 100% by AI agents, and this article would lose its title. 😄