Runs on this siteone afternoon · 70s render · 15s bake-off vs Synthesia

What I learned making my first AI explainer video with HeyGen

Primary tool: HeyGenBuilt with: Claude Opus 4.8, ChatGPT 5.6Stack: YouTubepricing as of jul 2026last verified jul 2026

Presenter and voice are AI-generated; scripted and directed by Jigg.AI.

Being AI-made isn’t a disclaimer here — it’s the whole point.

How I picked the tool (and why I didn’t just guess)

Two tools own this space: HeyGen and Synthesia. Instead of arguing about which was better, I ran a test. I wrote one 15-second script — deliberately booby-trapped with the things AI presenters fumble: my brand name said aloud (“Jigg.AI”), a spoken number (“thirty questions”), a clipped one-word beat (“No hype.”), and a fast three-item list — and fed the exact same text into both tools. Then I watched.

HeyGen read my script as written. Synthesia rewrote it — it turned my 15 seconds of first-person copy into 46 seconds of third-person corporate narration (“Our AI solutions undergo rigorous testing…”), retitled the whole thing “Jigg.AI: Practical AI Tools and Solutions,” and made a build-in-public project sound like a SaaS company that doesn’t exist.

Here’s the part I want to be fair about, because it’s easy to get wrong: that wasn’t Synthesia being broken. It was me leaving its AI-assist mode on, which is designed to expand a brief into a full presentation. I re-ran it with strict-script mode on, and it held my words at 20 seconds — right next to HeyGen’s. So the honest finding isn’t “Synthesia can’t follow a script.” It’s that the two tools have different instincts: HeyGen reaches for a creator/startup register, Synthesia for a presenter-led, training-video one. For a build-in-public site, HeyGen’s instinct fit my voice with less fighting. That’s the real reason I picked it — not a fake knockout.

The honest process — what broke

The free tier can't shipTool limit

HeyGen’s free plan watermarks the whole frame and caps you at 60 seconds. My explainer was longer, and a watermarked video isn’t publishable as the site’s featured build. So I paid. Expected, but worth knowing before you start: “free” means “evaluate,” not “publish.”

The tool invented my brand's design onceMy setup

On import it defaulted to a placeholder logo (“logoipsum”). Pulling my real brand kit fixed it in one step.

A template tried to credit a fake personMy setup

HeyGen’s “Name Card” template ships with a dummy “Lead Engineer” credit, and I couldn’t delete it. It never appeared — my script gave it no intro to hook onto — but on a brand with no named individual, that’s the kind of default you check for. Dodged.

It re-paced my scriptTool limit

I wrote ~87 seconds; it rendered 70. Faster, and nothing important got cut — but the tool re-times you, so verify your key line survived. (Mine did — I checked the honest-failure beat was still intact.)

The avatar I picked left black barsMy setup

On the plain talking-head shots, a portrait avatar sits in a vertical column inside the landscape frame. It reads as repurposed-vertical, and it’s the thing I’d fix first next time — by choosing a frame-filling avatar.

The test — how I’ll grade every video build

The 30-question format I use for bots doesn’t fit video. So here’s the video test, decided before I generated (not graded on a curve after): fresh-viewer clarity, avatar believability, voice naturalness, and the honest operational cost.

3/3
fresh viewers got the premise unprompted: 0
uncanny-valley flags on the paid render: 70s
finished video — no camera, no face

The clarity number is the one that counts most: it’s the thing I can’t self-judge, so I showed the video to people who’d never seen the site and asked what it’s about. All three got it. The “No hype.” beat and the list pacing landed, and “Jigg.AI” was pronounced cleanly.

HeyGen Creator plan, incl. tax: $32.06
bake-off that decided it vs Synthesia: 15s
v2 fixes logged: framing, proof, captions: 3

One honesty note on cost: $32.06 is the real paid Creator plan (with tax), which is what it took to ship a watermark-free video. It’s a monthly subscription, so a clean “cost per finished minute” depends on how much you make in the cycle — the concrete figure is the plan price. Time-to-render, number of takes and credits consumed weren’t logged during the session, so I’m not going to invent them.

Two AIs, arguing — how this build actually got made

This is the part I’d normally hide, so I’m showing it. This build wasn’t made by one AI. It was made by me pointing two of them at each other.

Claude did the planning and wrote the script. ChatGPT I used as the opposition — I had it argue the other side and tear up my drafts. They genuinely disagreed, and the disagreement made the build better:

On the tool. ChatGPT argued for Synthesia — better for a documented, reproducible, expandable video system, closer to the “real business tool” thesis. Claude pushed back that Synthesia’s “business” is enterprise training, not my reader, and that the prompt I’d publish lives upstream anyway. Neither won by argument. The 15-second bake-off settled it with evidence.
On the output. Once I had a render, ChatGPT critiqued it hard — flagged that the proof on screen was too small to read, that the avatar was boxed in black bars, and that one line (“an AI answer engine can’t fake”) overclaimed, since an engine can fake a claim; what it can’t fake is the visible evidence. Claude triaged those into must-fix vs nice-to-have and caught the one thing ChatGPT got wrong — its fix for the AI-presenter disclosure would have broken the brand’s first-person voice; the right fix was to disclose and keep the voice.

The takeaway, and the reason this section exists: the useful move wasn’t trusting either AI. It was making them disagree in front of me and keeping the judgment for myself. That’s the whole method of this site, applied to its own making.

Reproduce this

On this page: the honest process above, the scored selection rubric (HeyGen vs Synthesia, including the strict-mode fairness re-run), and the video test — all decided before the render was judged. The full rubric and results are in the repo at builds/explainer-video/test.md, and the running journal, fault-tagged, is in build-notes.md.

One honest note on provenance: HeyGen has no native way to export prompt history, so the prompts behind this build (research → script → scene-plan → revision) and the Claude↔ChatGPT exchange live upstream in the chats. They’re assembled by hand from the logs — not written from memory afterwards — so the pack lists what’s captured and marks what’s still being exported rather than filling gaps with plausible reconstructions. The pack is at builds/explainer-video/repro/.