One clip of yours and two photos become the same video with you in it, in another language or with other characters. Claude shows the plan and the price before spending anything.
This is a free INEMA kit (in Portuguese) with ready-made messages to paste into Claude Code, file templates, checklists and a price table. It teaches you to use Hypit, an open video tool, to redo a short clip with your own face, make a version in another language, or swap the person for fictional characters. It suits people who do marketing or create content and want to test several versions of an ad without filming again. You need Claude Code, a fal account with a few dollars of credit, a clip of yours and two photos of your face.
Hypit plans and assembles the video on your computer. fal runs the image and video models. Claude Code talks to you and writes the files. This kit gathers what you need to tell it.
Four ready-made messages to paste into Claude Code. No commands to type.
Every message ends with "Show me the plan and the price and don't build anything." Claude estimates the cost and stops.
After the first one, a new face or a new language is just one more message, not a day of filming.
Install, redo it with your face, check and make versions. In each one, Claude shows the plan and the price first.
One message installs Hypit just for the project, connects fal by reading FAL_KEY from the environment, and points to the Chrome you already have. About 6 minutes.
Hypit takes the first and last frame of the clip, puts your face in both, and generates the 8 seconds between them. Images first, video after.
Face, hands and text in the scene. Fix with premium models. Then: another language or fictional characters in the same clip.
Four things, plus Chrome. Claude checks the rest (such as Node.js 22.15+) and warns you if something is missing.
Runs on your Claude plan. Let it ask permission before running commands; no automatic mode for paid work.
A few dollars are enough to start. Create the API key in the fal.ai dashboard and keep it in your terminal profile:
# at the end of ~/.zshrc (Mac) or ~/.bashrc (Linux) export FAL_KEY="paste-your-key-here" # check without showing the value [ -n "$FAL_KEY" ] && echo "FAL_KEY loaded" # Windows (PowerShell): saves it for your user setx FAL_KEY "paste-your-key-here"
Never paste the key into the chat. Hypit reads it from your computer.
8 to 10 seconds (the messages use the first 8 s). Yours, or one you have the right to redo.
One facing the camera and one slightly from the side. No sunglasses, no hand over the mouth, with good light.
Paste each message into Claude Code, in the project folder. The full versions, with explanations, are in kit/ (in Portuguese).
An empty folder for the project, with the clip and the photos inside references/. You don't need to pull the frames out of the clip: Claude extracts the first and last frame when you send the step 3 message.
meu-clone/ ├── characters.json # only for step 7 (characters) └── references/ ├── meu-clipe.mp4 # the clip you are going to redo ├── rosto-1.jpg # photo facing the camera ├── rosto-2.jpg # photo slightly from the side └── meu-ugc.mp4 # only for step 7
Open Claude Code in that folder and paste the install message. What usually happens: Claude installs the skill only in the project (npx skills add hypit-ai/hypit, without -g) and the @hypit/hypit package, and sometimes fixes a setting that came out wrong on its own. You just approve. If /hypit is not recognized afterwards, close and reopen Claude Code in the same folder.
Install Hypit in this folder. Install the Hypit skill from github.com/hypit-ai/hypit for this project only, and the hypit command through npm. Then connect fal so the images and clips run on fal: Hypit doesn't come with fal, so write a small project Provider package in packages/provider-fal that reads my FAL_KEY from the environment and never writes the key to a file. FAL_KEY is already set. Point Hypit's page renderer to the Chrome I already have installed. Show me the plan first and then check that everything works.
First and last frame with your face, then the video between them. In two rounds, so you can check the images before paying for the video. If your clip is shorter than 8 s, replace "8 seconds" with the real duration. About US$0.71 by the 10/21/2026 price table.
/hypit Redo the first 8 seconds of references/meu-clipe.mp4 with my face, using references/rosto-1.jpg and references/rosto-2.jpg. Take the first and last frame of the clip and put my face in both, keeping the same pose, the same room and the same light. Then make the 8-second clip between those two frames. Split it into two rounds, images first and then the clip, so I can check the images before paying for the video. Show me the plan and the price and don't build anything.
The face (eyes, nose, smile), the hands (count the fingers in fast movements) and any text in the scene — a "CEREAL" sign can turn into "CENLIL". In versions with characters, also check hair and teeth.
A stronger image model, a middle frame and video in 1080p. Same prompts, only the model changes. About US$2.
/hypit The face and the text in the scene look fake. Redo it with premium models: GPT Image 2.5 at maximum quality for three frames (start, middle and end), so the movement doesn't go off track, and MiniMax H3 Max at 1080p for two clips of about 4 seconds each. Keep all the prompts the same, changing only the model. Show me the plan and the price and don't build anything.
Translates and trims the speech, new voice, check with Whisper, synced lips and subtitles. If the voice gets a number or a rare word wrong, write it the way it's spoken and check again. Use your own voice or a generic voice: cloning someone else's voice requires their permission. About US$2.
/hypit Make a Japanese version of my premium clip. Translate the speech and trim it to fit in 8 seconds, make a new voice, run a speech check, sync the lips with the new voice and add subtitles. Show me the plan and the price and don't build anything.
With the kit's characters.json (7 invented characters) in the root of the folder and the clip in references/meu-ugc.mp4. 7 versions cost about US$25.20. If the safety filter blocks a photo (content_policy_violation), soften that character's description and run only that one again.
/hypit I want the first 8 seconds of references/meu-ugc.mp4, a UGC clip, redone with different characters, one version per character in characters.json. Each character gets an identity photo made by GPT Image 2.5, and then MiniMax H3 Recast puts that person into the clip at 1080p, keeping the movements, the camera and the timing. Show me the plan and the total price of all versions and don't build anything.
Run them all as a single test, with the same small budget for each. After a few days, keep the two that the audience watches the longest and make new versions of the winner: face, setting or language.
Stay under 10 seconds. Recast accepts 5 to 30 s, but short gets it right more often and costs less.
No sunglasses, no hand over the mouth. One facing the camera, one from the side, good light.
Check the images (cents) before paying for the clip (dollars).
Costs calculated from the list prices read on 10/21/2026. Prices change: open the pages before spending. Hypit is free; you pay for the models on fal.
| What | Approximate cost | Source |
|---|---|---|
| First version (Seedream 5 Lite + MiniMax H3 768p) | US$0.71 | fal price table |
| Premium version (3 GPT Image 2.5 images + 2 H3 Max 1080p clips) | ~US$2 | fal price table |
| New language (voice, Whisper, lips, subtitles; Claude picks the models and shows them in the plan) | ~US$2 | fal price table |
| MiniMax H3 Recast 1080p | US$0.45/s → US$3.60 per 8 s clip; 7 versions (the characters.json) ≈ US$25.20; batch of 8 ≈ US$28.80 | fal · Recast |
| GPT Image 2.5 (editing) | ~US$0.006 to 0.055 per image + ~US$0.01 per reference | fal · GPT Image 2.5 |
| Higgsfield Genjutsu 1080p (comparison) | US$1.632/s → US$13.06 per clip; batch of 8 ≈ US$104.50 | Higgsfield |
Recast and Genjutsu are different models; quality was not compared here, only the list price. The "premium" and "new language" rows are approximate sums of the models cited: the exact value is what Claude shows in the plan.
A skill is a package of instructions that Claude Code loads. /hypit at the start of the message makes it use the Hypit skill.
A provider is the package that connects Hypit to a service (here, fal). The hypit.runtime.json says which model runs on which service.
Hypit's composition language: a file with the inputs, the prompts and the model jobs. Claude writes it for you.
AI models on fal: the first two make images; MiniMax H3 makes video. Recast is the H3 version that swaps the person in a video.
Whisper listens to audio and writes down what it heard (to check the new voice). Lip sync is syncing the mouth with the voice.
UGC is video that looks like it was recorded on a phone by a customer. 768p and 1080p are the height of the video in pixels: bigger means sharper and more expensive.
content_policy_violation is the response from the model's safety filter when it refuses a description. Node.js is the program that runs Hypit; Claude checks whether you have it.
Read the full license in the Hypit repository. That is why Hypit is not included in this kit.
MIT license: only INEMA's text, file templates and scripts. What you create with Hypit is yours; the model providers' terms still apply.
All in text, in Portuguese. Open it on GitHub or download the repository.
kit/ ├── 00-COMECE-AQUI.md ├── 01-instalacao/ what you need · install message · safe FAL_KEY ├── 02-mensagens/ the 4 messages for Claude ├── 03-modelos/ characters.json · example runtime · 2 .svml files ├── 04-checklists/ check before fixing · 3 mistakes · test with 7 versions └── 05-precos-e-links.md
Explainer video for the kit: video page.
