Feature
Writing for an avatar: nine rules and one read aloud test
Half of what people call a bad avatar is a good render of writing meant for the eye. The render did what it was told. The script told it the wrong thing. Here are the nine rules a script has to pass, the free test that catches almost every failure, and the length the credits and the attention both agree on.
- Words a minute the sync handles cleanly
- 150
- Practical ceiling before delivery costs you viewers
- 40s
- What the read-aloud test costs
- 0 CR
The nine rules, in the order they matter
Write it to be said, not to be read. A sentence with a semicolon is two sentences. Use two.
One idea per sentence. A subordinate clause forces the delivery to hold a shape it has no way of holding.
Aim near 150 words a minute. Above 180 the mouth shapes start overlapping. The sync then reads as late even when it is not.
Front-load the specific. A number in the first four seconds survives an even delivery. An adjective does not.
Cut the words that need emphasis to work. A rendered read is even. Anything relying on a stressed syllable loses its meaning.
Cut the questions you do not answer. A rhetorical question needs a delivery that leans on it and there is no lean available.
Name the product once, early, then stop naming it. The captions can carry it visually instead.
End on an instruction rather than a flourish. A flourish needs a performance to land.
None of these rules make the writing worse. They make it sound like speech, which is what the render is animating.
- No semicolons. Two sentences instead.
- One idea per sentence, no subordinate clauses.
- Near 150 words a minute, never above 180.
- A number or a name in the first four seconds.
- No line whose meaning depends on which word is stressed.
- No rhetorical questions. No sarcasm. No irony.
- No first-person feeling. No sensory claim. No apology.
- Say the product name once early and then stop saying it.
- End on an instruction, not on a flourish.
Words a minute, and what happens to the sync
The sync is derived from the waveform. Pace it beyond 180 and the mouth shapes start to overlap.
- Slow and deliberate120 wpm
- The reliable zone150 wpm
- The edge180 wpm
- Sync starts reading late210 wpm
The read-aloud test costs nothing and catches most of it
Stand up and read the script out loud at a normal pace. Once. Without stopping.
Every place you run out of breath is a sentence you built too long. The render will struggle with it.
Every place you naturally leaned on a word is a line that will flatten. The read has no lean in it.
Every place you felt awkward saying a claim is a line a viewer will not believe from a face that never felt it.
Mark those three and rewrite them before you spend a credit.
The test takes two minutes and is more accurate than any preview.
If you cannot say the script comfortably, no model is going to say it convincingly.
Then render twenty seconds of it as a test, for about 117 credits, before committing to the full length.
This is the whole editor
Highlight a phrase and a clip lands on those exact words. No timeline, no keyframes, no layers.
The lines to cut on sight
Anything beginning with I felt, I noticed or I could not believe. A generated face performing an unfelt reaction is the loudest tell in the format.
Anything with this in it while pointing. There is nothing to point at. The sentence leaves a hole where the demonstration should be.
Any sensory description. How it smells, how heavy it is, how the fabric moves.
Any sarcasm or irony. Both need a delivery the read cannot give. Both land as sincere instead.
Any sentence over roughly twenty-five words. The breath has to go somewhere and it will go in the wrong place.
Cutting these is not a compromise. It is sorting.
Keep the cut lines in a second document rather than throwing them away.
That document is your filming list, and it is worth more than the script you kept.
Write for forty seconds. The credits and the attention agree.
At 350 credits per rendered minute, forty seconds is about 233 credits. Ninety seconds is about 525.
Attention drops faster than the price rises. An even delivery has no small human variation to earn patience with.
So the ceiling is a performance decision. The budget happens to agree with it.
Cut the setup. A rendered presenter has no charm to spend on a warm-up. The first sentence has to be the point.
About 100 words is forty seconds at a comfortable 150 a minute. Write to the word count rather than to the clock.
Once the render is back, batch it for 100 credits. The editor covers up to 45 percent of the runtime with footage at normal pace.
Highlight a phrase and a clip lands over exactly those words. That is how forty seconds of script stops being forty seconds of face.
Emphasise a word so it pops in the captions. The caption is where the emphasis a read cannot give ends up.
Every other avatar tool leaves that step for you to solve somewhere else.
From script to a cut that holds
Two of these steps are free, and they are the two that decide the result.
Write to 40 seconds
About 100 words
Read it aloud
Free. Mark breath, lean, awkwardness
Rewrite those lines
Free
Render
About 233 CR at forty seconds
Batch and cover
100 CR, up to 45 percent footage
Where the lines you cut should go
A testimonial is testimony all the way down. Removing the feeling to make it renderable removes the reason it worked.
A demonstration script is a sequence of physical actions, and there are no hands in a rendered frame.
A founder note persuades because of who is saying it. Rendering it works against itself.
Those go on a phone. One take of up to three minutes batches for 100 credits and finishes in the same editor.
You direct it on the same transcript, cover the same phrases and export the same 9:16 file.
So the filming list is not a record of what this product cannot do. It is a record of what it does the other way.
Write the renderable half and the filmable half as separate documents from the start.
The plain facts to write around. Rendered audio caps at five minutes. A filmed take caps at three. Output is 9:16.
There is no timeline, by design. Every lever is a decision about the ad.
Both documents end up in the same account and come back as the same finished file.
Questions people ask
- How many words is a forty second script?
- About 100 at a comfortable 150 words a minute. If your draft is 140 words, you are writing at 210 a minute and the sync will read late. Cut, do not speed up.
- Can I write in a conversational style with filler words?
- A little helps and a lot hurts. One so or right at a sentence start reads naturally. Repeated filler exposes the even delivery. A real person varies pace around filler and a read does not.
- Should the script mention the brand name often?
- Once, early, then stop. Repeated naming works when a person's warmth carries it and grates when the delivery is even. The captions can carry the name visually instead.
- Which scripts should not be written for an avatar?
- Testimonials, demonstrations and founder notes. Film those on a phone for 100 credits a batch and keep the render for announcements and explanations.
Read it out loud once, cut to a hundred words, then let one batch cover half the runtime with footage. No script fixes that last part on its own.