DEV Community

Cover image for Clone yourself as a character, then let Claude Code make the explainer videos
Syed Abdul Hanan for Logit Solutions AI

Posted on AI-assisted

Clone yourself as a character, then let Claude Code make the explainer videos

The host in this clip was drawn from one photo.

A photo of the Mona Lisa, an arrow, and the cartoon host drawn from it, waving and saying: I was drawn from one photo, and now I can host your explainer.

It blinks, waves, points and moves its lips to the words, and it can present an explainer on any topic you ask Claude for. The photo here is the Mona Lisa, a public domain painting, standing in for yours.

We built this as selfclone, a free, open-source Claude Code skill. Version 0.2.0 is out now. Here is what it does, how it works and where it falls short.

One photo in, a host out

selfclone character --photo me.jpg --consent "This is my own face. Your Name" --name "Your Name"
Enter fullscreen mode Exit fullscreen mode

The skill finds the landmarks of your face with MediaPipe, takes the proportions from them and samples the colours from the photo. Out comes a drawn SVG character: a rig with jointed arms, a face with six moods, and mouth shapes for the sounds it speaks. You can pick the hair (twelve styles), facial hair and outfit. It also writes a preview sheet like this one:

The cartoon character drawn from the Mona Lisa in eight poses: rest, smile, open mouth, rounded mouth, serious, pointing, waving and explaining.

It is a cartoon on purpose, not a deepfake of you. It refuses to run without a consent statement, and it stores that statement inside the character file.

Then ask for a video

Ask Claude for "an explainer about compound interest, hosted by my character". Claude checks the facts, writes short sentences with one idea per scene, and lays the video out as a scene file. This is the first scene of one of our examples:

{
  "title": "How does money grow?",
  "cast": { "host": { "character": "random:money-5", "voice": "kokoro:af_sarah" } },
  "music": "ambient",
  "scenes": [
    {
      "say": [
        "Put one thousand dollars away at seven percent a year.",
        "After one year, you have one thousand and seventy."
      ],
      "background": "#8fc6a0",
      "backdrop": "dots",
      "title": "Compound interest",
      "character": { "x": 0.2, "y": 1.0, "scale": 0.85, "enter": "rise" },
      "items": [
        { "prop": "coin", "x": 0.58, "y": 0.45, "w": 0.12, "at": 1.0, "loop": "bob" },
        { "text": "${n}", "count": { "from": 1000, "to": 1070, "dur": 1.6 }, "x": 0.78, "y": 0.45, "size": 0.12, "line": 1 }
      ]
    }
  ]
}
Enter fullscreen mode Exit fullscreen mode

Before rendering, selfclone explainer scenes.json --preview writes three stills per scene and prints storyboard warnings: items that overlap, run off the frame, sit under the captions or cover a character, and scenes that are empty or rushed. Claude fixes the file, then renders:

selfclone explainer scenes.json --character .selfclone/character.json
selfclone explainer scenes.json --aspect 9:16    # vertical, for Shorts, Reels and TikTok
selfclone explainer scenes.json --html           # an interactive player with chapters and captions
Enter fullscreen mode Exit fullscreen mode

Each render writes the MP4, an .srt caption file and a provenance file that records how the video was made.

What happens during a render

  • Voice. Kokoro, an 82M-parameter text-to-speech model, runs on your CPU. It is Apache-2.0, pinned and hash-checked. ElevenLabs, OpenAI or your own recording also work.
  • Mouth. The lip sync is read from the text: closed lips on M, B and P, teeth on the lip for F and V, rounded O and U, timed to the voice.
  • Body. Seven gestures (point, wave, explain, present, shrug, think, rest), chosen from the words when the script gives no cues: a wave on "hello", a shrug at a question.
  • Scene. 60 hand-drawn props, bar, line and pie charts that grow, handwriting that writes itself, and Manim for maths.
  • Frames. Headless Chrome draws every frame, with the network blocked, across parallel workers.

There is no GPU and no image model in the loop. One of the examples below, a 37-second explainer about the seasons, rendered its 900 frames in 27 seconds on six workers, on a laptop.

Ten explainers, ten fields

A showreel of ten animated explainers made with selfclone: neural networks, how planes fly, compound interest, vaccines, the Pythagorean theorem, who or whom, neurons, the Ship of Theseus, why the sky is blue and the seasons.

Every one of these came out of the same skill, with a different cast each time: one host, two characters in dialogue, or nobody on screen and a narrator. They use random fictional people (random:<seed>), so no real face appears. The topics are neural networks, how planes fly, compound interest, vaccines, the Pythagorean theorem, who or whom, neurons, the Ship of Theseus, why the sky is blue (as a vertical short) and the seasons.

The scene files are in the repo under examples/showcase/, so you can read exactly how each one was written. The v0.2.0 release has the showreel with sound.

What it is not

  • Not hand animation. Bodies move in 2D, a head turn is a shift of the face, and walking is a simple cycle.
  • Not photoreal. It does not try to be.
  • Not a fact checker. A video is only as good as its scene file. Read the script before you render; the preview stills are there to catch layout mistakes.

It also clones how you write

The other half of selfclone is a writing clone. Give it your own chat exports and it builds a persona skill that replies the way you write. Then it scores the clone on replies it never saw, by counting style features instead of asking a model to grade itself. On the fictional example in the repo, the clone scores an 89% style match, against 41% for a bare model. The README has the method and the limits.

Consent and labels

  • A real face needs its owner's consent, and the consent record travels inside the character file.
  • The photo avatar, selfclone's other face feature, animates the photo itself, so it always carries an "AI CLONE" label that cannot be switched off.
  • Explainer videos carry their provenance in the file's metadata and in a .provenance.json next to the video.

Try it

In Claude Code:

/plugin marketplace add logitsolutionsai/selfclone
/plugin install selfclone@selfclone
Enter fullscreen mode Exit fullscreen mode

Then ask: "make me a character from me.jpg and an explainer video about how vaccines work". You need Node 20 or newer and ffmpeg. Manim is optional, for maths. The skill also works in Codex and Cursor.

It is MIT-licensed and on GitHub: logitsolutionsai/selfclone.

If you had a cartoon version of yourself, what would you have it explain first?

Top comments (0)