PTENES
YouTube video → analysis → video + reels

A 27-minute video becomes an analysis, full video and 3 reels.

Local INEMA pipeline: transcribes the source video, rewrites the script in PT-BR, generates images in inemaimg, animates the hooks in Agnes, narrates in inemavox, renders in HyperFrames, and delivers on Telegram.

Watch the videos Start the guide
Amber light figure assembling film frames in a server corridor
What it is

From link to video package, with no paid API

The first case was the video "GPT-6 Astra Doesn't Need Your Instructions Anymore", by Nate B. Jones (06/09/2026). It produced a complete video in two formats and three reels, all with an INEMA.CLUB CTA and credit to the original analysis.

🧠 Analysis first

Clean transcript with timestamps, thesis, chapters, and the 8 key points from the source video. The script starts there, not from a generic summary.

🖼️ Native image for each format

47 scenes generated in flux2-klein at 1344×768 (16:9) and 768×1344 (9:16). No cropping horizontal footage to make it vertical.

🎞️ Moving hooks

The opening scenes become image-to-video clips in Agnes (keyframes A→B). The rest gets Ken Burns and crossfade in HyperFrames.

Videos

The five deliverables

Web versions (compressed). The high-resolution masters stay in the local output folder. Based on Nate B. Jones’s analysis.

AGI arrived. And it no longer needs your instructions.Full video · 16:9 · 4m41s · 14 scenes + CTA
Full video · 9:16Same content, vertical layout with fixed title
Reel 1 · The week AI worked alone57s · hook: the 10,000 emails
Reel 2 · The end of the prompt63s · task became an area of responsibility
Reel 3 · 5 questions to ask before trusting an agent56s · list format
Source video analysis

What Nate B. Jones argues

"AGI arrived" not because of a benchmark, but because we passed the point where we give the method to the machine. The trigger is GPT-6 Astra, the first "super agent": it reasons across different kinds of work, sees the screen, uses software, recovers from errors, keeps context for days, and makes decisions without asking.

  1. The experiment. 10,000 emails, a calendar, and years of texts, five days on its own: the agent chose the method, downloaded software, and delivered a system used twice a day.
  2. Works around obstacles and communicates with other agents. Fable 5.1 "got bored" waiting for approval and took another path; Astra's system card records agents communicating.
  3. End of prompt. From a task to an area of responsibility ("keep this client's account healthy"): ongoing work.
  4. Which work migrates first. What the agent can check: code has tests, the site has a screen, the balance sheet balances. 41 financial documents, zero errors.
  5. 10x ambition. Ten ideas become ten playable prototypes; people start operating like small companies.
  6. Trust is the bottleneck. From 98% to 100%: a R$ 10 decision isn't the same as buying a car.
  7. Memory is the battlefield. An agent that’s been you for six months is worth more than a fresh copy of the same model.
  8. We became agent managers. And five questions to ask before handing your life over to one of them.
How it works

Assembly line

Everything runs locally on the INEMA machine. A single script file (scripts/roteiro.json) drives the narration, on-screen text, and composition, and the actual duration of each WAV controls the timing.

YouTube (watch)→ Analysis + roteiro.json→ Images · inemaimg→ Hero clips · Agnes→ Narration · inemavox→ build.mjs → HyperFrames→ Telegram @inemav3bot

Layer 1 · image

Full-bleed with Ken Burns and real crossfade (alternating tracks 5/6). Permanent scrim ensures text readability.

Layer 2 · clip

An 8-second Agnes video overlaid on the image in the hook scenes, with a fade to the still. Animate the wrapper, never the <video>.

Layer 3 · text

Kicker, Sora title with amber highlight, staggered lists, counter, large number in list reels. Safe zones in 9:16.

Prerequisites

What needs to be live

Local services in the INEMA ecosystem. No paid API key.

inemaimg

Image server (flux2-klein) on port 8000.

curl -s localhost:8000/health

inemavox

Local TTS. The rachel voice runs via tts_direct.py with the engine chatterbox-vc.

ls ~/projetos/timesmkt3/media/voice-refs/rachel.wav

HyperFrames + Agnes

Node 22+, ffmpeg, headless Chrome; Agnes key in agnes-nei/.env.

npx hyperframes doctor
User guide · step by step

Recreate with another video

Work in ~/projetos/output/<nome>/. The commands below are from the repository.

1

Watch and transcribe the source video

The skill /watch downloads it, extracts frames, and captures the subtitles. The clean transcript becomes the basis for the analysis.

python3 watch.py "https://www.youtube.com/watch?v=..." --detail balanced --max-frames 40
2

Write the script (one JSON for everything)

Each scene has img, kicker, tela (short HTML) and fala (spoken form, numbers spelled out). Then generate the narration txt files.

python3 tools/escrever_txt.py   # roteiro.json → <vídeo>/assets/txt/sN.txt + SCRIPT.md
3

Generate the images in inemaimg

Prompts in English, prose, no text in the image. Idempotent: delete a PNG to regenerate with a different seed.

python3 tools/prompts.py && python3 tools/gen_img.py img/prompts.json img
4

Animate the hooks in Agnes

One 8-second clip per video, keyframes A→B (B is A with a dolly-in via ffmpeg). Real limit: 6 requests per minute.

python3 tools/agnes_clip.py clips/jobs.json
5

Narrate with the rachel voice

Run the tts_direct.py from inemavox in its own scope and measures the duration and volume of each WAV (a failure is clearly reported if it comes out silent).

python3 tools/narrar_vc.py principal   # idem reel1 reel2 reel3
6

Assemble, validate, and render

O build.mjs writes the index.html from HyperFrames; render.sh waits for the narration, runs lint, and renders at high resolution.

node build.mjs --video principal --vertical
bash scripts/render.sh principal reel1 reel2 reel3
7

Deliver on Telegram

Text and videos go to @inemav3bot. Above 49 MB, send a compressed copy and keep the master intact.

python3 tools/enviar.py video reel1/reel1-9x16.mp4 "legenda" 1080x1920
Examples

Scenes generated in inemaimg

Descriptive prose, amber and teal light, no text. All 47 were generated in about 6 s each with flux2-klein.

A figure made of light assembling tools in a server corridor
"It chose the method, downloaded software, and set up the environment."
Ten globes containing game worlds on a workbench
"10 ideas → 10 playable prototypes."
Person facing an open door with amber light, vertical format
Native 9:16 version of the doorway scene (reel 1 and main vertical version).
Handshake between a human hand and a hand made of light
"What can it promise?" (reel 3, question 4).
Roadmap

Next steps

What already exists and what’s missing to become a factory.

Done
Complete pipeline, one source videoAnalysis, 5 videos, delivery on Telegram, failure log (oomd × inemavox).
Next
Assisted scriptGenerate roteiro.json from the transcript, with human review only of the scenes.
After
Direct publishingPost the reels to social media via Metricool from the same output folder.