You provide the question ("why does the fridge heat up the kitchen?"). The AI researches it with sources, writes 4 hooks, shows you the storyboard, and only then records the voice and assembles the video. Two agents manage the account and ads without spending anything without your OK. You don't need a graphics card: without one, the reel comes out with captions and music.

curiosidade-viral is a kit (a "skill" for Claude Code or Codex) that makes vertical curiosity reels without anyone appearing on camera: a voice explains while the screen shows an everyday object up close. It’s for anyone who wants to create a short-form content account without filming or editing. The kit checks the hook before you spend, asks for your approval on a storyboard, records the voice on your computer, and assembles the video with code. It also includes two agents that read messages and ad metrics and only make recommendations. You need Node.js, ffmpeg, Python, and Claude Code or Codex.
In a real one-week experiment, an account that started from scratch had a reel with more than 4 million views and reached about 19,000 followers. What made the difference afterward was the model: what to offer, how to measure, and who decides.
Lead with the strange, make it personal, leave a question, make it instantly understandable, and make it work without sound. Write 4 and compare them in pairs.
One sheet per scene with a sketch, lines, timing, and what you see. Voice, images, and rendering only start after you say "go ahead".
Curio sorts the messages and prepares replies. Clic reads the ads and suggests cutting or scaling up, with the 4 rules written in code.
The key piece is the storyboard. Nothing costs anything before it; after it, everything else is a script.
The hook’s object appears large in almost every scene. First as people know it, then from the inside.
Heat, force, current: drawn over the real thing in a single color. What’s wrong in red. Numbers with the source in gray.
The camera pulls back until the reel becomes a phone, and a finger taps "Send". The caption stays at the top, away from the platform buttons.
Nothing here calls a paid API. The voice runs on your computer.
Rendering uses HyperFrames (it downloads automatically the first time).
node --version ffmpeg -version
For the checker, storyboard, agents, and music in code. On newer Ubuntu/Debian systems, pip system package is blocked: use the system package.
sudo apt install python3-numpy # on Mac: python3 -m pip install numpy
Writes facts, hooks, the script, and scenes. It’s what "uses" the skill: the instructions the kit gives the AI.
claude # or: codex
O voz.py uses Chatterbox (clones your voice from 10–20 s of audio) and Whisper (timestamps each word), running on an NVIDIA graphics card. Specify their locations with the variables below. The voice worker comes from inemavox (free): clone it and follow its README. No GPU? Skip the voice: the reel will have captions and music (montar.py --sem-voz).
export CHATTERBOX_PYTHON=~/miniconda3/envs/chatterbox/bin/python3 export INEMAVOX_DIR=~/inemavox # folder with venv and Whisper: git clone https://github.com/inematds/inemavox export VC_VOZ=~/my-voice.wav # required: YOUR voice (otherwise it looks for the author’s default voice)
Run these from the kit folder. For steps 3, 4, 8, and 9, ask Claude Code or Codex; the rest are commands.
Enable the skill in Claude Code and Codex, run the tests, and create your preferencias.md (audience and closing).
git clone https://github.com/inematds/curiosidade-viral.git cd curiosidade-viral bash instalar.sh
A short name and the question. It comes out exemplos/<nome>/ with roteiro.json, fatos.md and pie.txt to fill in.
python3 ferramentas/novo.py chuveiro "why does the shower go cold when someone turns on the faucet?" --duracao 30 --idioma en
Ask the AI: "fill in fatos.md with sources." For each number, include its value, the exact sentence, the link, and the date. Leave out anything without a source.
# in Claude Code, inside the kit folder: make the fatos.md for exemplos/chuveiro following the SKILL.md
The AI writes 4 hooks, each in a different style (contrarian, object, number, personal, question), the curiosity-loop script, and the scenes. The rules are in metodo/gancho.md and metodo/roteiro.md.
write the 4 hooks and the script for exemplos/chuveiro
Flags openings with a date or "according to a study," a news tone, "I/my channel," hooks with the same format, numbers without a source, excessive duration, missing closing, or hashtags in the post.
python3 ferramentas/gancho_check.py exemplos/chuveiro # OK = 0 errors
Right after step 2, it reports several errors (everything is empty): that’s expected. Run it again after step 4.
The only key piece. One card per scene with a sketch, lines (key phrase highlighted), timing, entrance, and sound; at the top, the hook and 3 backups. Approve it, change a scene, or swap the hook.
python3 ferramentas/storyboard.py exemplos/chuveiro --abrir
Records each line and marks the second for each word, so the screen changes on the right word.
python3 ferramentas/voz.py --projeto exemplos/chuveiro exemplos/chuveiro/roteiro.json --simular # only shows what it will record python3 ferramentas/voz.py --projeto exemplos/chuveiro exemplos/chuveiro/roteiro.json
3 to 5 photos of the SAME thing: take them with your phone (from the front, open, from behind, close up), in neutral light and without text, and put them in exemplos/chuveiro/media/. If you have Codex, you can ask it to generate the photos. The AI builds the reel page by copying the fridge example.
build the scenes for exemplos/chuveiro based on exemplos/geladeira
Choose one of the two paths. Before rendering (about 5 minutes), capture still frames and look at each scene.
# A) WITH recorded voice in step 7 (GPU) python3 ferramentas/montar.py --projeto exemplos/chuveiro python3 ferramentas/mezcla.py --projeto exemplos/chuveiro --id chuveiro --musica tranquila python3 ferramentas/preparar.py --projeto exemplos/chuveiro
# B) WITHOUT voice (no GPU): captions + music; the duration comes from the text python3 ferramentas/montar.py --projeto exemplos/chuveiro --sem-voz python3 ferramentas/mezcla.py --projeto exemplos/chuveiro --id chuveiro --musica tranquila python3 ferramentas/preparar.py --projeto exemplos/chuveiro
# then, in both cases cd exemplos/chuveiro npx hyperframes check npx hyperframes snapshot --at 0,5,12,20 -o stills # check the images in stills/ npx hyperframes render --output renders/chuveiro.mp4 cd ../.. # go back to the kit folder
One link per reel, so you know what brought each person in. Post from your phone: in the real-world case, the automated post tanked reach.
python3 agentes/link.py https://your-site.com --conta myaccount --onde reel --reel chuveiro
The rules are in the code, not just in the prompt: the agent can’t forget any of them. Cards to paste into your agent: agentes/curio.md and agentes/clic.md.
Separate DMs and comments: "how do I do it?" gets a ready-made reply with the link; complaints, refunds, and brand mentions go to you; like compliments. Only notify you when there's news. The script doesn't send anything: you copy and send it, or connect the form agentes/curio.md to one of your agents.
python3 agentes/curio_triagem.py agentes/exemplos/mensagens.json \ --link "https://your-site.com/?utm_medium=dm" --estado curio.json
1. no OK, no spend · 2. daily cap · 3. only scale up with a return of 1.5x or more · 4. quickly cut what doesn’t sell. And two alerts: broken tracking and scattered budget.
python3 agentes/clic_regras.py agentes/exemplos/metricas.jsonThe agents don’t access your account: you (or the tool you already use) export the data to two simple files. The kit includes examples of both in agentes/exemplos/.
New DMs and comments, copied by hand or exported from the platform.
[{"id": "d1", "de": "@marina", "onde": "dm",
"texto": "How do you make these videos?"}]From the ad manager (spend, budget) and your store (sales, revenue), in the same currency.
{"config": {"teto_dia": 150, "retorno_min": 1.5},
"medicao_ok": true, "compradores_excluidos": true,
"anuncios": [{"id": "geladeira-h1", "orcamento_dia": 40,
"gasto": 82, "vendas": 4, "receita": 178}]}# Clic output with the kit example (the tools respond in Portuguese)
Clic · proposta de anúncios (nada muda sem o seu OK)
⚠ Quem já comprou ainda vê o anúncio: exclua os compradores do público.
SUBIR geladeira-h1: 40 → 50/dia · retorno 2.17 ≥ 1.5 com 4 venda(s)
CORTAR geladeira-h2: 30 → 0/dia · gastou 35 sem nenhuma venda (limite 30)
↻ geladeira-h2: subir um anúncio novo com outro gancho da reserva no lugar
REDUZIR cebola-h1: 30 → 15/dia · vende, mas devolve menos do que custa (retorno 0.72)
Total proposto: 85/dia (teto 150).
The full example is at exemplos/geladeira/: sourced facts, 4 hooks and the selection, script, storyboard, voice, images, and the reel page. To make your own, copy this folder and replace the content.

Numbers from a one-week experiment, not a promise of results.
The question I got most often was "how do you do that?". That’s exactly what you offered.
On a day without tracking, 19 people opened the payment page and no one bought. Without tracking, nothing scales.
A few ads with a budget returned about 2 times the spend; many ads with a small budget returned about half.
Sales you can’t trace back to their source can’t be repeated.
The agents respond and keep watch. You decide how to spend the money.
The bio brought in 7 sales without a single ad.
A folder of instructions that Claude Code or Codex follows to complete a task. Here, to make the reel.
The reel’s first sentence and first image: what makes someone stop scrolling.
The reel laid out on cards, scene by scene, before it exists.
The reel’s short folder name, with no spaces or accents (e.g., chuveiro).
The text for the post caption.
Each sentence answers the previous one and opens a new question, until the final answer.
The tool that turns an animated web page into an MP4 video.
For every 10 spent on the ad, 15 came back in sales.
The same link with an ending (?utm_…, the link tracking pattern) that tells you where the person came from: which reel, the bio, or a DM.
Direct message, sent privately on the social network.
The 3 hooks that weren't chosen: use them for a B version of the same reel or for a new ad.
The maximum the ads can spend per day, combined, in the currency you use in metricas.json.
Free programs that run on your computer: Chatterbox speaks in your voice; Whisper listens and marks the timing of each word.