I stopped filming my reels (my AI avatar reads the script)

AI avatar reels sound like a gimmick until you count what filming actually costs you: the setup, the lighting, the 14 takes, the day you didn't post because you didn't feel like being on camera. The talking head is the one part of a reel that can't be batched. Unless it isn't you.

This is the pipeline I use now: a script goes in as a text file, Claude Code sends it to HeyGen's API, my avatar reads it in my cloned voice, and the clip lands in the folder my editor pipeline already watches. I still write every word. I just don't film them.

What you'll have when you're done

• An avatar of you and a clone of your voice, reusable by ID
• A skill file that turns a script file into a 9:16 talking-head clip with one sentence
• A cost and length check you approve before anything renders (every render is real money)
• Three hook variants and the body in one render, ready for the trial-reel test
• [SCREENSHOT: script.md on the left, the finished avatar clip on the right]

Before you start

• Claude Code installed (it's Claude in a terminal that can run tools on your machine). Comes with the Claude Pro plan, $20/month.
• A HeyGen account. Creating the avatar and voice happens in the app; the API runs on a prepaid wallet, about $0.05 per second of finished video for a photo avatar. Load $10 and you have three or four reels of runway.
• One good photo of you (chest up, even light, plain background, looking at the lens) and 1 to 2 minutes of clean audio of you talking.
• 30 minutes.

Step 1: Make your avatar and clone your voice

This part happens once, in the HeyGen app, and takes about 10 minutes.

1. Avatars, then create a photo avatar from your photo. Give it a name you'll recognise later ("Studio hoodie", "Desk, glasses"). HeyGen calls each generated variation a look; you'll want the ID of the look you like.
2. Voices, then clone your voice from your audio. Name it after yourself.
3. Approve the consent step for your own likeness when HeyGen asks. It's required, and it's the right rule.

Start with a photo avatar. It's ready in minutes and it's the cheapest per second. If you later want the one that moves more like you, train a digital twin from a couple of minutes of footage and swap one ID in the skill file.

Step 2: Give Claude Code the key and let it find your IDs

Open Claude Code in a new folder (call it avatar). Get an API key from HeyGen's settings and put it in a file called .env in the folder, one line, HEYGEN_API_KEY=your-key. Claude Code reads it from there; you never paste it into a prompt.

Now let it fetch your IDs instead of hunting through the app:

Read HEYGEN_API_KEY from the .env file in this folder (never print it).
Using HeyGen's v3 API (base https://api.heygen.com, header X-Api-Key):
list my avatar groups with GET /v3/avatars, then the looks in each with
GET /v3/avatars/{group_id}/looks, and my private voices with
GET /v3/voices. Show me a short table: name, type, ID. Then check
GET /v3/users/me and tell me my wallet balance.

Pick the look ID and the voice ID you want. Keep the table; the skill file uses both.

Step 3: Write the avatar skill file

A skill file is a saved instruction document Claude Code reads before working, the same idea as the editor's EDITOR.md. Create it once, with your two IDs filled in:

Create a file called AVATAR.md in this folder with exactly this content
(replace the two IDs with mine from the table), then confirm you'll follow
it whenever I ask for a render:

You are my avatar renderer. When I give you a script file, run this:

1. READ the script. It has four sections: # HOOK A, # HOOK B, # HOOK C,
   # BODY. Strip any [SCREEN: ...] cue markers and markdown emphasis.
   Join the sections in that order with a blank line between each.
   Never add SSML or break tags; anything in angle brackets gets read
   aloud.
2. SHOW me the final text, the estimated length (words divided by 2.6
   is seconds) and the estimated cost at $0.05 per second, then WAIT
   for my approval. Every render costs money.
3. RENDER: POST https://api.heygen.com/v3/videos with header X-Api-Key
   from HEYGEN_API_KEY and this JSON body:
   type "avatar", avatar_id "<LOOK_ID>", voice_id "<VOICE_ID>",
   script <the text>, aspect_ratio "9:16", resolution "1080p",
   fit "cover", engine {"type": "avatar_iv"}, expressiveness "medium",
   caption {"file_format": "srt"}, motion_prompt "calm talking-head
   delivery, small head movements, light hand gestures at chest level,
   stays centered". Save the video_id from the response.
4. POLL GET https://api.heygen.com/v3/videos/<video_id> every 15
   seconds until status is completed. If it fails, show me the error
   and stop. Download video_url to renders/<scriptname>.mp4 and the
   subtitle file next to it.
5. REPORT the file path, the duration, and the balance before and after.

Rules:
- Never render the same script twice unless I say "re-render".
- Never print the API key.
- Never edit my script. If a line reads wrong out loud, tell me.

The approval gate in step 2 is the point of the whole file. A render is a purchase; you want to see the text and the price before it happens, every time.

Step 4: Write a script, render the first clip

Write your reel as a text file, one sentence per line, hooks first:

# HOOK A
I stopped filming my reels. You're literally watching my AI avatar right now.
# HOOK B
Okay, this isn't actually me. This is my AI avatar, and it made this whole reel from a text file.
# HOOK C
What if you never had to film another reel? Yeah, you're looking at the answer.
# BODY
[SCREEN: script.md being typed] Here's how it works. I write the script in a text file. Honestly, that's the only part I still do.
[SCREEN: the skill file] Then Claude Code reads a little skill file. That's where my avatar, my cloned voice, and my rules live.
...

Three hooks in one render costs a few cents more than one and gives you three trial-reel variants for free. Then:

Read AVATAR.md, then render script.md.

Claude Code shows you the text, the length, and the price. Say "approved" and go get coffee. My first render (three hooks plus the body, 58 seconds) went from pending to completed in about 3 minutes; the clip appears in renders/, and the balance report tells you exactly what it cost. Two things I learned on that run: 4K is accepted for photo avatars but bills at about $0.067 per second instead of $0.05, and Whisper hears "HeyGen" as "Hey Jen", so expect one caption fix downstream.

Step 5: Hand it to the editor, then batch it

The clip is raw footage, and it goes through exactly the same editor pipeline as filmed footage: transcribe, cut, caption, screen recordings, export. If you built the editor from the previous guide, drop the clip in that folder and run it. The one difference: no retakes to clean up, so the cut list is short.

Then batch. Write five scripts on Monday, render five clips in a row, edit them the same afternoon. The camera never came out.