Hark's AI - I cloned my voice for done videos

This is Hark's AI, my voice clone by ElevenLabs & Claude Code.

What you need

• An ElevenLabs(opens in a new tab) account on the Starter plan ($6 a month). It is the cheapest plan with Instant Voice Cloning.
• Claude Code(opens in a new tab)
• A few minutes of you talking, with no music behind it. (You can also use your already published YouTube videos)

Step 1 - Create an API key

In ElevenLabs, go to Developers > API Keys > Create Key.

Name it done-videos, so future you knows what it is for.

Give it only what it needs, or your SysAdmins will get mad at you:

Only the first 2 are required.

Everything else can stay on No Access.

Keep Auto-disable if leaked on.

Step 2 - Put the key on your machine

Paste this into Claude Code:

Save my ElevenLabs API key as an environment variable called ELEVENLABS_API_KEY,
in a file that stays out of git. Do not ask me for the key.
Open the file and I will paste it in myself.

I keep mine in ~/.zshrc.local. Paste the key, save, restart the terminal. Then ask Claude to:

Check ELEVENLABS_API_KEY is set and works. Do not print the key.

Never paste the key into the chat.

Step 3 - Clone your voice

Clone my voice with ElevenLabs Instant Voice Cloning from these videos: <your YouTube links>.
Use only the parts where I am talking alone, with no music.
Name the voice "<your name>'s AI".

Or clone it in the ElevenLabs web app: Voices > Create Voice > Instant Voice Clone.

Step 4 - Use it

Make a 2-3 minute done video for <PBI link>, narrated with my ElevenLabs voice "<your name>'s AI".
Write the script, voice it with eleven_v4, burn in subtitles, and build the video with Remotion.

That is the prompt I started from. Claude Code wrote the script, made the narration and built the video.

⚠️ Notes from my runs

AI voices are not deterministic. My 2nd video came out sounding American 😅

So I measured what keeps it close, and these are the things I follow now:

• Tell it your accent. I start every request with [Indian English accent]. In a side-by-side test, the same scene went from 100% American to 0% with just this tag.
• Keep stability at 0.5 on eleven_v4. At 0.7 it sounded calmer, but the clone stopped listening to the accent tag.
• Send 500-900 characters per request, and pass the previous request's ID (previous_request_ids) so the tone carries over. ElevenLabs says very short requests drift more.
• Fix the seed (the random roll behind each take, like "seed": 1729). Same text and same seed gave back a 99% identical take, so I can redo one line without the voice changing.
• Slow it down with pauses, not the speed slider. eleven_v4 ignores speed.
• Do not redo a scene for a mispronounced name. Mine says "Clod" for Claude sometimes. The subtitles say Claude, and redoing it just burns credits.
• Do not pay for sound effects or music. Royalty-free clicks, typing and background music from the YouTube Audio Library work fine.

With all of these together, across a whole 3 minute video:

I packed these into a Claude Code skill for my own done videos. It is on GitHub(opens in a new tab) if you want to see how it works.

What it costs

Starter gives 30k credits a month, about 1 credit per character on eleven_v4. A 3 minute script is about 3k characters, so with a few redos that is around 8 videos a month. Unused credits roll over for 2 months.

2 rules I do not skip

• Every video says it is my AI, in the intro and the sign-off. On YouTube, I tick "altered or synthetic content".
• Only clone your own voice, or one you have permission for.