AI avatar video without a camera — a checklist

This is the version to keep open while you work. You open it during production, move phase by phase from the face clone through to the edit, and check off each step. Every item is something you can verify — you run it and confirm it works before moving on.

Three tools, one chain: HeyGen builds the avatar, ElevenLabs (11 Labs) builds the voice clone, Remotion stitches everything together and adds captions and animation. Claude Code operates the chain: it reads files, runs the steps, and talks to the other tools. You say what you want; it carries it out.

Phase 0 — before you start

• Have a finished script ready in a file on disk — that's still your thinking, the machine only generates the face and the voice.
• Set up a HeyGen account and confirm you have access to Avatar 5; the older versions, Avatar 3 and 4, give themselves away with stiff mouth movement and mechanical gestures.
• Set up an ElevenLabs account — this is where the voice comes from, since HeyGen's default voice sounds weak.
• Set Claude Code up as the tool that chains the steps, so you're not copy-pasting script chunks or downloading and uploading recordings by hand.
• Total the costs before you start (rough figures, prices change — check current rates):

Phase 1 — face clone in HeyGen

• Pick a route for the avatar recording:
• Faster — a webcam recording: read the short text the tool displays, record for about fifteen seconds, get the finished clone back a moment later.
• More precise — upload your own footage: at around ten gigabytes of recordings, the model captures your expressions better.
• Generate in Avatar 5 — it's trained on more than ten million data points on facial expression and builds a double from a fifteen-second recording that turns its head, blinks, and swallows.
• Check the result: the figure should look natural enough that it's hard to tell from a real shot. If something is "off" about the mouth, you're probably working on an older version.

Phase 2 — voice clone in ElevenLabs

• Do the voice separately in ElevenLabs — this is the step that matters most for how the final piece lands.
• Choose a voice-clone mode:
• Instant voice clone — fast and simple, but less faithful.
• Professional voice clone — needs material: at least thirty minutes of recordings; around two hours gives the best result.
• Adjust pace, stability, similarity, and style strength in the panel — expect several iterations before it lands on your way of speaking.
• Export the finished recording as an audio file.
• Upload that file to HeyGen in the AI Studio section and sync it with the avatar.
• Don't import the same voice straight into HeyGen — it sounds worse there than in ElevenLabs. That's why you build the voice separately.

Phase 3 — cutting up the script

• Work within two tool limits:
• HeyGen on Avatar 5 generates clips of up to three minutes.
• ElevenLabs starts degrading the voice after about a minute, and it drifts further from the original.
• Cut the script into chunks of forty-five to sixty seconds — the voice stays consistent throughout at that length.
• Hand the cutting to Claude Code: it takes the text from disk, splits it into roughly one-minute pieces, and passes them to HeyGen.
• Make sure every cut lands at the end of a sentence, never in the middle — that's what keeps the seams inaudible once the pieces are stitched together.
• Work around the fact that Avatar 5 isn't in the API yet: Claude Code generates through Avatar 4 via the API, and a separate Playwright script (it clicks around the browser like a person) opens the HeyGen panel and switches the version to five. This workaround is temporary — it goes away once Avatar 5 reaches the API.

Phase 4 — editing in Remotion

• Give Remotion the background and the visual style.
• Let it assemble the recordings on its own.
• Check the transcription — Remotion converts speech to text with exact timestamps.
• Confirm the captions and animations land on the right second: since the tool knows a word fell at second forty-four, it fires the animation exactly there, so the audio lines up with the graphics.
• Kick off the whole chain in the evening and pick up the finished video in the morning.

What to watch for

• The voice degrades after a minute. Don't push segments past about a minute in ElevenLabs; beyond that it drifts further from the original.
• Cut only at the end of a sentence. A mid-sentence cut is audible as a seam once the clips are joined.
• Build the voice separately, never straight into HeyGen. Importing the same voice directly into HeyGen sounds worse than in ElevenLabs.
• Avatar 5 isn't in the API yet. While only Avatar 3 and 4 are available there, you need the Playwright workaround; drop it once Avatar 5 reaches the API.
• This doesn't work on the first click. Behind a finished result sit more than a hundred, maybe two hundred generated clips and a fair number of settings attempts — expect several rounds of fixes.
• Prices are approximate. Tool rates change — check current numbers before you budget.

---

The bottleneck here shifts away from recording and editing and onto what the machine won't do: the idea, the script, and the strategy. The tools compress production, but a good result still needs your grasp of the subject and patience for the fixes.