Skip to content
Carnero logo
◉Email Carnero

How it works

A content engine with approval gates at every step

The Carnero engine takes a format and a topic all the way to a published, measured post. Dozens of parameters shape every video, every fact is checked before it can be used, and configurable rules decide what runs automatically and what waits for approval.

The pipeline

From a format to a measured post

  1. 01

    Format and topic

    A skill (a reusable format) and a topic from its queue, written in or suggested by AI and approved.

  2. 02

    Verified facts

    The topic is ready only when its facts have passed the fact gate.

  3. 03

    Plan and script

    An AI text model writes the plan: the title, every shot, the exact words to be spoken and the music direction.

  4. 04

    Approval and cost hold

    The plan passes an approval gate, directly or through its approved calendar slot, and each paid step is priced and held first.

  5. 05

    Production

    AI video scenes, narration, music, captions with the spoken word highlighted, and the logo.

  6. 06

    Checks

    Opening, upright-picture and character checks, narration length, and technical checks for black frames, freezes and silence.

  7. 07

    Publishing

    Texts for each network within its limits, the AI disclosure, and the upload with the planned visibility.

  8. 08

    Measurement

    Results from YouTube and Meta, a weekly report, and recommendations for the next videos.

Capabilities

One engine, from the idea to the measured result

Planning, production, fact checking, publishing, measurement and cost control in one system, with approval gates at every step.

Languages

Narration in English and Hebrew; captions and dubbing in six languages.

  • Narration in production in English and Hebrew.
  • Captions burned into the video with the spoken word highlighted, in the video’s language or translated into English, Hebrew, Spanish, Arabic, French or German. Hebrew and Arabic are drawn right to left.
  • Multi-language dubbing of a finished video into English, Hebrew, Spanish, Arabic, French or German. The dubs published so far are in Spanish, on YouTube, Instagram and Facebook.

Content planning

Reusable formats, topic queues, AI topic suggestions and a weekly calendar.

  • Skills: reusable video formats with their structure, writing and visual rules, and title patterns.
  • A topic queue per format. The engine suggests topics with AI; each suggestion is queued or rejected at an approval gate.
  • Series tags on topics, such as “X vs Y”.
  • A weekly calendar built from a template, with slots per day and network that pass an approval gate.

Quality and truth

Facts from approved sites with exact quotes, plus picture, opening and sound checks.

  • The fact gate: every fact needs an approved site, a page the engine reads itself, and an exact quote with its numbers.
  • An opening check of the first two seconds, a check that each scene is upright, and a check that recurring characters stay consistent.
  • A narration length check: narration that is too short for the film is rewritten and spoken again.
  • Technical checks on every finished video for black frames, frozen picture and silence.

Sources and fact policy

Publishing

YouTube, Instagram and Facebook, with texts written for each network.

  • Uploads to YouTube, Instagram Reels (also shown in the feed) and a Facebook Page.
  • TikTok: the connection is built; until TikTok’s app review is approved, TikTok allows private posts only.
  • Titles, descriptions and hashtags for each network, suggested by AI on request and marked as AI suggestions.
  • Every post carries the AI disclosure sentence.

Measurement and learning

Results from YouTube and Meta, a weekly report, and recommendations.

  • Results are collected from YouTube, Instagram and Facebook one hour, one day, three days and seven days after publishing.
  • A weekly report in the app, every Sunday.
  • Recommendations from the results, comparing length, hour, title style and opening. Applying them is a deliberate choice, never automatic by default.

Cost control

A ceiling per video, daily and monthly caps, and a hold before every paid call.

  • A spending ceiling per video, and daily and monthly caps per profile.
  • A weekly budget for the calendar.
  • Before any paid AI call, the engine prices it from a recorded tariff and holds the money. With no current price, or over a cap, the call is not sent.

Control and safety

Approval gates, separate profiles, encrypted keys and nightly backups.

  • Approval gates for video plans and calendar slots, with configurable rules for what runs automatically and what waits for approval.
  • Profiles, each with its own accounts, settings and limits.
  • Sign-in tokens and service keys stored encrypted.
  • A verified nightly backup.

Under the hood

The parameters behind every video

Set per format, profile, calendar slot or single video, depending on the parameter, and applied at every step from the script to the upload.

Format and structure

  • Shape: vertical 9:16, wide 16:9 or square 1:1.
  • Length within a range set per format, for example 15 to 35 seconds for Shorts.
  • A structure of an opening, three to five scenes, a payoff and a loop line, with shots of 4 to 15 seconds.

Language and voice

  • The language of each format, and a narration voice for each language.
  • A voice for each dubbing language.
  • Narration pace in words per second. Once a voice has been measured enough, the measured pace replaces the setting.
  • Narration fill: how much of the film the narration covers, from tight (90 to 95%) to roomy (55 to 70%).

Opening

  • An opening of one to two seconds, in a chosen style: a comparison, a question or a search-style line.
  • Warnings for silence before the first word, too few words in the first two seconds, and a long first scene.
  • A picture check of the opening frames, and examples of the strongest and weakest earlier openings in the series given to the planner.

Facts and numbers

  • The script may state a number only if it is in the topic’s checked facts.
  • Numbers and on-screen text are burned in by the engine, not drawn by the video model.
  • Every fact passes the fact gate before it can be used.

Picture

  • A visual style and a list of things to avoid, per format.
  • Recurring characters from reference images, checked scene by scene.
  • Every scene checked for being upright; a sideways scene is re-created once, then held for review.

Captions

  • The spoken word highlighted, with the coming words dimmed.
  • A maximum line length, and font size and margins for each shape.
  • Outline, shadow, and fade, lead and hold timing. On or off for each video.

Sound

  • A music style for each format.
  • Music lowered under the narration, with fades in and out.
  • Loudness normalized to −14 LUFS with a −1.5 dBTP peak, after the mix.

Texts for each network

  • A title pattern, a description footer and base tags for each format.
  • Each network’s limits: a YouTube title of up to 100 characters and up to 50 tags, an Instagram caption with up to 30 hashtags, and a Facebook title of up to 255 characters.
  • A target number of tags or hashtags for each network, with room kept for the AI disclosure.

Timing and visibility

  • Time zone, default publish time and a daily upload limit for each channel.
  • Private, unlisted or public visibility.
  • Calendar slots for each day and network, with lead times before publishing.

Budget

  • A ceiling for each video, and ceilings for single AI calls such as topic suggestions and post texts.
  • Daily and monthly caps, a weekly calendar budget and a maximum for each slot.
  • Limits on attempts, rewrites and scene re-creations.

Quality gates

  • Opening, character and upright-scene checks.
  • Narration length: too-short narration is rewritten and spoken again.
  • Black frames, frozen picture and silence, measured on every finished video.

Results

  • Results collected from YouTube, Instagram and Facebook at one hour, one day, three days and seven days.
  • Comparisons by length, hour, title style, opening and template, once each group has at least five videos.

In production

What the engine has made so far

Counts from the engine’s own records, measured 17 October 2026.

finished videos, all passed the technical checks
64finished videos, all passed the technical checks
narrated videos uploaded to YouTube
47narrated videos uploaded to YouTube
song videos and song Shorts uploaded to YouTube
15song videos and song Shorts uploaded to YouTube
dubbed videos published to YouTube, three of them also to Instagram and Facebook
5dubbed videos published to YouTube, three of them also to Instagram and Facebook
topics in the planning queues
87topics in the planning queues
video plans reviewed at the approval gate
80video plans reviewed at the approval gate

AI services

The models behind the engine

Each service receives what its step needs, for example a prompt, a script, a text to speak, or a picture to check.

MiniMax
Video scenes.
ElevenLabs
Narration voices, songs, and speech timing.
Google Lyria
Scene music.
Google Gemini
Text, such as plans and research, and picture checks.
OpenAI
Text, images, picture checks, and a built-in narration voice.

How the engine handles data is described in the Privacy Policy. Privacy Policy