Builder Créatif × Créer Société Dubaï
Two raw clips.
Edited and live in minutes.
🎥 raw-01.mp4 · 4:12🎥 raw-02.mp4 · 2:47
Here's the raw footage. Get me the final cut and publish it on my platforms?
Got it. On it:
- Analyzing the footage
- Word-level subtitles
- Hook burned into the first 3 seconds
- “Uhs” and hesitations removed
- Branded intro + outro
- Publishing to TikTok · Instagram · YouTube
Live on all 3 platforms, each with its own caption. Links below 👇
A real request, condensed into a few seconds.
You film thirty seconds of yourself between two meetings — false starts, hesitations, your eyes drifting to your notes and all. You send the raw video to the Telegram bot. What comes back is edited in your brand's identity: the hook burned into the first three seconds, subtitles timed to the word, the excess “ums” taken out. Live on six platforms, each with its own caption. It's the editing seat that disappears — not your face.
- Telegram
- 2 clips → 1 edit
- Word-perfect subtitles
- Per-network captions
- Autonomous publishing
What follows isn't a product promise: it's the blueprint of a machine, station by station, with its guardrails.
The economics
Cost stops tracking
volume.
This is the only number that really matters. With an editor or an agency, doubling the videos doubles the invoice — the curve is a rising straight line. With the machine, the bill is a roughly fixed monthly infra cost: going from 14 to 100 videos a week moves it by a few euros.
Move the sliders. The red line is what you pay today. The blue line is the machine.
The script is translated and re-voiced: each language is one more
video, for the same amount of human work — none.
What you're comparing
What the benchmarks are worth. €120 per video is the
low end of a Dubai editing retainer (AED 5,500/month for 8 to 12
videos). €350 is an agency-produced video (AED 700 to 3,500). On the
machine side: €400 of infra per month, plus roughly €1 of render and
tokens per video. I took the low end on the human side and the high
end on the machine side — the real gap is wider than the one shown,
and I would rather quote infra slightly above the real cost than
below it.
⚠️ That €461 is what the machine costs, not my price. What it costs me to
run is not what you pay me: my work is the build (the 4 weeks) and then running it. That is
priced against scope — we discuss it live, and I give you a number, not a range.
The problem
The bottleneck was never
the idea for the video.
You know exactly what to say. Freezone versus mainland, what a visa really costs, what people get wrong about tax — you answer these questions on the phone all day. The content is already in your head.
What stalls is everything between the idea and the post: writing the script, finding an editor, waiting three days, reviewing, fixing a number, re-exporting, then pasting the same caption onto six networks and hoping it lands everywhere. That journey costs more than the idea, and it's why you publish once a month instead of three times a week.
An agent doesn't replace your expertise. It deletes the journey.
- Idea noted somewhere, never picked up again
- Brief to the editor, back-and-forth over WhatsApp
- 3 to 10 days between the idea and going live
- One caption copy-pasted everywhere
- Manual publishing, network by network
- Two raw clips sent inside Telegram
- Clean-up, hook, subtitles and captions right away
- Live the same day
- Six captions, six platform playbooks
- Automatic publishing and reporting
The machine
The full loop,
station by station
Six stations. Only one is human — the first, and it's all we ask of you. The arc along the bottom is what makes the machine interesting over time: how each video performs, and every correction you make, flows back into the agent's brain and shapes the next brief.
Hover a station to isolate it. Drag the diagram sideways.
The learning curve
It starts seasoned.
It never plateaus.
An editor joins at their level and stays there: their ceiling is their experience. The agent arrives already seasoned — trained on short-form codes and your brand rules. And every published video pushes it up: the hooks that retain, the cuts that hold attention, the captions that convert — everything is measured, then fed back into the next brief.
A few dozen videos in, it knows your audience better than any contractor ever will — because it holds the exact memory of everything that worked and failed on your accounts.
Station 1 — the interface
No dashboard.
A conversation.
The best tool is the one you already open fifty times a day. The agent lives in Telegram: you write to it like a colleague, voice notes included — it transcribes, it understands, it replies. No software to learn, no extra tab, and it works from the back of a taxi.
🎥 2 clips · 4:12 + 2:47 · unedited
“Make a video about the number one mistake French founders make setting up in Dubai: they pick a freezone for the tax, when their activity actually requires mainland. Keep the tone direct.”
Got it. 41s → 34s. 7 of 11 hesitations removed — I kept the ones that breathe. Hook burned in: “4 in 10 clients arrive with the wrong structure.”
🎬 video-freezone-vs-mainland.mp4 · 47 s
Video edited, 6 captions generated. Charter-compliant — publishing now.
Live on all six networks. Here are the links:
Hook is too soft, be more direct
Written into the charter. New version in 4 min.
Stations 2 and 3 — the edit
The video isn't edited.
It's compiled.
This is the reversal that makes everything else possible. Instead of dragging layers around in an editing suite, the video is described in code: a template declares where the logo sits, how the title enters the frame, what colour the subtitles take. The agent doesn't “edit” — it fills a template and triggers a render.
The direct consequence: the hundredth video is as clean as the first, the brand identity never drifts, and restyling every future video means editing one file.
What the template guarantees
- The logo never moves. Same position, same size, same safe margin, on every single video.
- Colours are variables. Your azure and your gold are declared once. No video can ship off-brand.
- Subtitles are generated, not typed. The transcription is aligned to the audio word by word.
- Every format is a variant. The same content renders in 9:16, 1:1 and 16:9 without re-editing.
Remotion
Video written in React- Every frame is a React component — so it lives in Git, versioned.
- Mature ecosystem: render on AWS Lambda, hundreds of videos in parallel.
- The template becomes a real product you evolve instead of rebuilding.
The sensible default as soon as you're after volume and longevity.
See the project →HyperFrames
Video written in HTML- Renders from HTML/CSS: the agent writes what it already knows how to write.
- Very short authoring loop — ideal for iterating on a look.
- Built-in cloud rendering, less infrastructure to run on day one.
Excellent for finding the style fast, before freezing a template.
See the project →My call: start on HyperFrames to nail the look in a few days, then freeze the winning template on Remotion once the cadence picks up. Both render MP4 from code — moving between them throws nothing away.
The technical sidebar
Should you render in 4K?
Almost always no — and that's good news for the bill. A vertical social video is watched on a phone, and the platform re-encodes whatever you upload, every time. Rendering in 4K means quadrupling the pixels for a result the viewer won't distinguish in their feed.
| Criterion | 1080 × 1920 recommended | 4K — 2160 × 3840 |
|---|---|---|
| Weight (60 s) | 15 to 25 MB | 60 to 100 MB |
| Render time | baseline ×1 | ×3 to ×4 (4× the pixels) |
| On an older machine | no trouble at all | runs hot, crawls, can fail |
| Once posted | re-encoded by the platform | re-encoded too — at the same bitrate |
| Cloud render cost | a few cents | ×3 to ×4 |
When you plan to recrop or punch in after the fact, reuse the master on a screen in an office or at an expo, or archive an interview that will keep earning for years. There, the extra pixels genuinely work.
A local 4K render on a Mac that's a few years old means fans at full tilt, a machine you can't use during the export, and renders that fail at the finish line. The fix isn't buying hardware: it's rendering in the cloud. The machine only fires off the job.
The master is produced once in 1080 vertical, then derived into 1:1 and 16:9 from the same template. What drives a short video's performance isn't its resolution — it's the first three seconds.
The edit
This isn't a filter.
It's an editor making calls.
Any tool can chop a video up. What you actually pay an editor for is judgement: what to cut, what to cover, when to post. That is exactly what the agent takes over.
- 01 It cuts the excess “ums” — not all of them Bursts of hesitation, false starts and dead air go. The one that breathes stays. This isn't fussiness: across 1,380 recorded calls, perfectly fluent delivery wins agreement just 3% of the time, against 36% for speech that hesitates a little. Aiming for zero “um” is aiming at the wrong number. Conrad et al., JRSS-A, 2013
- 02 Your glance at your notes becomes an image When your eyes drift off camera, it doesn't make you reshoot: it lays a chart, a diagram, a screen capture or B-roll over those seconds — your audio keeps running underneath. The flaw becomes the illustration.
- 03 You shoot five videos back to back, it publishes one a day Send them all in one session. It edits them, queues them and spaces them out at the right slot. You show up every day across the six networks without having to film every day — and on social, it's regularity that pays, not intensity.
Station 4 — distribution
Same video.
Never the same caption.
This is the most common and most expensive mistake: posting the same text everywhere. Each network has its own mechanics — LinkedIn cuts at “see more”, TikTok judges within the first second, YouTube behaves like a search engine. The agent starts from one script and writes six different texts, each in its platform's codes.
Hook One sharp line, then a line break right before the “see more” cut.
Tone Hard numbers, no decorative hashtags, expert-advisor voice.
Hook A direct question addressed to “you” in the first second.
Tone Warm, story-driven, 8-12 hashtags at the end of the caption.
Hook A divisive statement. You enter mid-story, never at the start.
Tone Spoken, blunt, oversized subtitles, CTA in a pinned comment.
Hook A searchable title — a Short behaves like a search query.
Tone Keywords in the title, long description, chapter markers.
Hook Local context first: “If you're an expat in Dubai…”
Tone More explanatory, full sentences, link at the end of the post.
Hook The strongest number, alone, on the first line.
Tone Optional thread under the video. Dry, factual, no emoji.
Bonus the second invoice
Writing captions and publishing
is its own line item.
Editing is only half the spend. Writing six captions, scheduling them for the right slot, posting to six accounts and pulling the numbers back — in Dubai that's bought separately, between AED 3,000 and 15,000 a month. It's exactly the part station 4 and station 6 absorb, with no extra line on the infra bill.
| Service in Dubai | AED / month | ≈ € / month | What it covers |
|---|---|---|---|
| Basic absorbed by the agent | 3,000 – 5,000 | €700 – 1,175 | 12-16 posts, captions, scheduling |
| Growth | 5,000 – 8,000 | €1,175 – 1,880 | 20-25 posts, reels, community management |
| In-house team | 15,000 – 30,000 | €3,530 – 7,060 | 1 to 2 people + tools |
What the agent does
- Write one caption per platform, in its codes
- Adapt format and cropping to each network
- Schedule and publish across the six accounts
- Pull back views, retention and engagement
- Send a report into Telegram
What it doesn't
- Reply to comments and DMs
- Handle a crisis or an unhappy client
- Decide the quarter's editorial strategy
- Run paid campaigns
The honest math. The agent replaces the "Basic" tier — captions, scheduling, publishing: roughly €700 to €1,175 a month, or €8,400 to €14,100 a year, on top of the editing saving. It does not replace community management or strategy: replying to a prospect in the DMs is still your job — and that's exactly where you want to spend the time the machine gives back.
The guardrails
“Autonomous” doesn't mean
“out of control”
The agent publishes without asking — that's the whole point: a tool that demands your sign-off on every video adds a task to your day instead of removing one. But publishing fast on your company's behalf, in a field where one imprecise sentence about tax or timelines puts your credibility on the line, has to be bounded. The brake isn't your inbox: it's the charter upstream, and reversibility downstream.
- 01 It publishes on its own — and nothing is irreversible There is no gate to clear: from your raw video, the machine runs all the way to publication. Your control doesn't sit in front of every video — it sits upstream in the charter, and downstream in the report: the agent sends you the links to all six posts. If something is off, you delete the post, you tell it why, it regenerates and the rule goes into the charter. On the early drafts you'll correct more often. That's normal, and each correction is only paid for once.
- 02 It only speaks in your words The agent reads an editorial charter file: approved phrasings, forbidden promises, the figures it is allowed to quote. In a regulated field like company formation, that file is what stops it from inventing a deadline or a guarantee.
- 03 An audit trail for every action Every video keeps its script, its template, its render version, the publish time and the account used. A post that causes a problem can be found — and pulled — in a minute.
- 04 A narrow blast radius, by design The agent has access to a content folder and to publishing accounts. That's it. No client CRM, no mailbox, no personal data. What it cannot touch, it cannot break.
- 05 A kill switch One command in Telegram pauses everything: no renders, no scheduled publications. An agent you cannot stop isn't a tool, it's a liability.
Trust gets built the other way round: the agent publishes from day one. In the early weeks you'll pull and correct more often — and every correction goes into the charter, permanently. You're not sliding an autonomy dial: the agent simply gets fewer things wrong, because your taste becomes a written rule.
What changes
The real win isn't cost.
It's cadence.
An agency delivers beautiful videos. It doesn't deliver the ability to shoot one the same evening because a client asked an interesting question at 5pm. The machine makes that reflex possible — and on social, consistency beats perfection every time.
The brief. Dictated. That's all the machine asks of you as input.
One script, six formats, six captions written for their platform.
Instead of a week of back-and-forth. The topic is still hot.
The template doesn't drift. Video one hundred looks like video one.
Under the hood
What I actually
build
- Interface
- Telegram — private bot, text and voice notes, automatic transcription
- Agent
- A scoped autonomous AI agent: persistent memory, declared skills, action log
- Video render
- Remotion (React → MP4) or HyperFrames (HTML → MP4), template in the brand's colours
- Subtitles
- Word-level transcription aligned to the audio, burned in at render time
- Publishing
- Multi-network APIs — LinkedIn, Instagram, TikTok, YouTube, Facebook, X
- Feedback
- Views, retention and engagement pulled back in and fed into the next brief
None of this is a demo prototype. It's the agent I already run daily to drive my own apps and my own content, on Telegram, with these guardrails. What this page describes is that engine, wired to your brand and your accounts.
The method
One meeting a week.
The rest runs without you.
Four weeks, and it runs on its own. The build runs autonomously: I'm not asking you to follow a worksite, sign off mockups or review specs. Between two check-ins, the agent and I move on our own. There is exactly one fixed appointment — one hour a week.
And that hour isn't a status update. It's there to extract from you the one thing no agent knows how to produce.
- Week 0 Kick-off
We pull out the soul and the goal
The only session that starts from a blank page. We brainstorm: what you want to say, to whom, and what you refuse to promise even if it would perform. We lock the angles, the tone, the red lines. That is exactly the material an agent cannot invent on its own.
- Written editorial charter
- The first 15 angles
- Template in your brand identity
- ↓ the agent builds — autonomously, without you
- Week 1 MVP
The first render
Not a mockup, not a slide: a real edited, publishable video, plus the Telegram bot answering. You watch it and you say what rings false. It is the single most useful moment of the whole project.
- 1 complete video
- 6 captions
- The bot replies in Telegram
- ↓ the agent builds — autonomously, without you
- Week 2 Phase 1
Your taste becomes a rule
Every correction you make isn't lost: it gets written into the charter. “That, never” turns into a constraint the agent can no longer cross. Week after week, it gets fewer things wrong.
- Template stabilised
- Tone rules encoded
- Cadence held
- ↓ the agent builds — autonomously, without you
- Week 3 Phase 2
Volume and languages
We wire in the six networks, scheduled publishing at the right slot, then the languages. This is where the curve from the top of the page pays off: volume climbs, the infra bill doesn't move.
- Multi-network publishing
- Multilingual
- Automatic reporting
- ↓ the agent builds — autonomously, without you
- Day 30 Turnkey
It runs on its own
One month after kick-off, the machine runs without me. You send your video, it ships to six networks. All that's left is the infra — €400 a month — and it keeps getting better on its own: every correction and every metric sharpens the next one.
- Turnkey solution
- €400 / month of infra
- It improves over time
What the meeting goes looking for An agent knows how to produce.
An agent knows how to produce.
It doesn't know how to want.
It writes fast, edits fast, publishes without slipping. But it has no idea what is right for your brand — because taste isn't inferred from a corpus, it's decided. Vision is the same: knowing which topic matters now comes from your client calls, not from a model.
That's why the weekly check-in isn't a progress meeting. It's the moment we take what's in your head and write it into the charter — so the agent can apply it for the next six days.
- What rings true — and what rings false, before you can even explain why
- The angle we keep and the one we bin, when both would “work”
- What you refuse to say, even if it would rack up views
- The topic that deserves to be next, because you heard it three times on the phone this week
Where this lands
The AI isn't an assistant.
It does the job itself, end to end.
It doesn't wait on you, doesn't ask what to do, doesn't hand back a draft to fix. You gave it the taste, the charter and the guardrails once — after that it produces, it publishes, and you discover the result at the same time as your audience. Without having spent any time on it.
That's the destination, not day 1. At the start you approve everything, and autonomy is earned format by format — exactly what station 5 describes.
The objections
The questions you're asking yourself
before saying yes.
Everyone asks them. Here they are, unfiltered — with the real answers.
Will people see it's edited by an AI?
It's your face, your words, your footage. The AI generates nothing: it edits, subtitles and publishes to your brand rules — your colors, your validated hooks, your cutting rhythm. The output looks like an editor applying a strict brief, because that's exactly what it does. And if a video doesn't feel like you: delete it, say why, the rule enters the charter.
What if the agent publishes something stupid in my name?
It starts in copilot mode: nothing goes out without your OK in Telegram. Autonomy is earned step by step, and even fully autonomous you keep the delete + learn button. The guardrails above aren't decorative: banned topics listed, validated tone, zero improvisation on figures or regulation.
How much of my time does it actually take per week?
The time to shoot two clips — 15 to 20 minutes, from your phone, between two meetings. No written brief, no editing back-and-forth, no content calendar to fill: you send, the machine does the rest.
Why not just a freelance editor or an agency?
Go back to the sliders in the economics section: with a contractor, doubling volume doubles the invoice, and every correction costs a 48-hour round trip. Here, the cost is near-fixed infrastructure, the machine runs overnight, and a correction = one Telegram message. An editor remains a great option — for people publishing two videos a month.
Who owns the machine — and what if we stop?
You do. The infrastructure runs on your accounts, the social logins stay with you, and the charter is a readable document that belongs to you. If we stop, everything stays in place: the system, the history, the content. No dependence on an opaque platform or yet another subscription.
Does it work in French, with an accent, in my industry?
Yes. Subtitles are word-level accurate in French and English, and the charter encodes your industry: vocabulary, sensitive topics (tax, immigration), mandatory disclaimers. That's exactly where the agent pulls ahead — it doesn't recycle generic templates, it applies your rules.
Next
We start with one video,
not with a contract.
The right way to judge this machine is to watch it run on your topic. Give me an angle you know inside out — freezone versus mainland, the real first-year budget — and I'll come back with the video edited in your brand identity and the six captions. You judge the output, not the pitch.