analisevideo — what a filmmaker sees in your video

What it is

Audiovisual language analysis, not transcription

The transcript already exists (inemaVOX/Whisper). What was missing was someone to look at the video and explain how it was made.

🎥 Camera, block by block

Each clip has a timecode: shot, angle, movement (dolly, gimbal, whip-pan, drone, orbital), speed, and what it does in the scene.

🎨 Lighting, color, and editing

Hex palette, lighting setup, look, and apparent lens; cuts per minute, pacing, transitions, cuts on the beat, and use of slow motion.

🗂️ Searchable database

Every analysis becomes JSON + a report and is added to a local index. Then you can search drone, search synthwave, stats.

How it works

From link to report in one call

Everything runs locally except the visual analysis, which is handled by Gemini.

link or file→ yt-dlp (≤480p)→ ffprobe: duration and format→ Gemini: a filmmaker's perspective→ analise.json + analise.md→ index.jsonl (database)

Why 480p

Color, lighting, and movement are easy to read at this resolution, and uploads are fast. Above 18 MB, the file is re-encoded to 360p/12fps for analysis only.

What it doesn’t do

Doesn’t run Whisper, return captions, or clone music. For speech, use inemaVOX; for cloning a track, use the musica skill.

Command only

In the bot, runs only on /analisevideo. A standalone link in Telegram still goes to the transcript, by design.

Prerequisites

What needs to be on the machine

No server: just shell + python3 + a Gemini key.

Binaries

Download, metadata reading, and database lookup.

# Ubuntu/Debian
sudo apt install ffmpeg jq python3
pip install -U yt-dlp

Gemini key

Loaded at runtime from .env from OpenPCBot (or wifi). It’s never copied here.

# ~/projetos/openpcbotv2/.env
GOOGLE_API_KEY=...

Clone

The repo is the skill: OpenPCBot points here via symlink.

git clone git@github.com:inematds/analisevideo.git \
  ~/projetos/analisevideo

User guide · step by step

Analyze, read, and query

Real commands. S=~/projetos/analisevideo/analisevideo.sh

Analyze a link

Downloads, reads the visuals, and archives. The report path appears on the last line.

bash $S analisa "https://www.youtube.com/shorts/XXXX"  # downloads, analyzes, indexes

Analyze a local file

Works for material yt-dlp can’t fetch (logged-in site) or for a clip that’s already been cut.

bash $S analisa ~/videos/reel.mp4 reel-marca --keep-src  # saves fonte.mp4

Focus the analysis on a question

--prompt adds a specific request, answered in the field pedido_extra.

bash $S analisa "<url>" --prompt "como eles seguram a atenção nos 3 primeiros segundos"

Read the report again

Markdown ready to paste into a brief, script, or work order.

bash $S ver reel-marca      # report
bash $S json reel-marca     # structured, for scripts

Query the database

This is where it becomes a library: search by movement, look, soundtrack genre, or tag.

bash $S list 10               # last 10 analyses
bash $S search "drone"       # everything with drones
bash $S stats                # most frequent movements and tags

On Telegram, through OpenPCBot

The skill is enabled as a bot command — it doesn't trigger on its own from a link.

/analisevideo <url|caminho>   # summary + analise.md attached
/analisevideo search drone
/analisevideo stats

Examples

Excerpts from a real report

Output from a 5s analysis of an animated infographic.

Cinematography

Paleta: #5DA9E9, #E85D3F, #46729B,
        #8C9BAE, #58B868, #FFFFFF
Luz: iluminação digital uniforme
Cor: equilibrada · contraste médio
Look: infográfico didático / flat 3D

Camera

00:00–00:05 · geral · estático
  · plongée — quadro fixo com o
  diagrama animado por camadas

Database

$ bash $S search tornado
teste-agnes [motion graphics explicativo]
  tags: infografico, motion graphics,
  tornado, meteorologia, didatico

Roadmap

Where it is and where it’s going

The core works today; everything else builds on it as a library.

Done

Analysis + databaseDownload, cinematic analysis, Markdown report, searchable index, Gemini retry, and bot command.

Compare videosPut two analyses side by side (pacing, palette, hook) to choose a project reference.

After

From the database to productionTurn the field reproduzir into a straightforward work order for pixflow / Kling / Veo.