Synthesia Video Workflow: Beginner to Pro Production Guide

Why Workflow Discipline Beats Feature Knowledge

Any modern avatar video tool can produce a talking presenter in under ten minutes. That is exactly why so much of the output looks the same: a generic face, a flat delivery, a wall of bullet points, and a runtime the script never deserved. The limiting factor in synthetic video production stopped being access to technology a long time ago. It is now the sequence in which you make decisions.

Teams that consistently publish watchable avatar videos follow a similar order of operations. They define the job of the video before touching an editor. They write the narration for the ear rather than the eye. They cast the presenter deliberately. They build a reusable visual kit instead of restyling every project from scratch. They review in structured passes rather than reactively. And they plan repurposing before the first render, not after.

Skip any one step and the cost shows up somewhere: a rushed narration, a scene that cannot be localized, a brand that looks slightly different in every episode, or a revision cycle that takes four times longer than the original build.

This guide covers the full arc: planning, scripting, casting, voice design, scene construction, editing, localization, export, and the mistakes that quietly consume the most hours. It is written for people who already know where the buttons are and want output that looks deliberate.

Define the Job, the Constraints, and the Project Structure

Write the one-sentence objective

Before opening the tool, complete this sentence in writing: after watching, the viewer will do something specific. Approve a policy. Troubleshoot a login failure. Understand why the roadmap shifted. Book a demo. If you cannot fill the blank with a verb, you do not yet have a video, you have a topic.

The objective determines everything downstream. A compliance module needs certainty and repetition. A sales outreach clip needs warmth and brevity. A product walkthrough needs screen footage and almost no face time. One objective per video. If you find two, you have two videos.

Pick four numbers and hold them

Write these down and treat them as a budget:

Target runtime.
Scene count.
Maximum spoken words per scene.
Deadline, including review time.

Synthetic narration runs at roughly 130 to 150 spoken words per minute depending on the voice and the pace. A 90-second explainer therefore holds around 200 to 225 words of narration, not 400. Teams that skip this arithmetic end up compressing a five-minute script into two minutes, then blame the voice for sounding breathless.

A workable default for a two-minute explainer: eight to twelve scenes, 15 to 25 words of narration per scene, one visual idea per scene.

Separate evergreen from dated content

Classify the video before writing. Evergreen content, such as onboarding basics, product fundamentals, or safety practices, should avoid anything that expires. Dated content, such as a quarterly update or a launch announcement, should be built so that the facts which change live in a small number of isolated scenes. Re-rendering a scene is trivial. Rebuilding a project because a number is buried inside an animated graphic is not.

Establish a naming and version system early

Version chaos begins with files labelled final-v3-use-this. Adopt a pattern such as project-topic-language-version-date and apply it to projects, exports, scripts, and transcripts. Add a simple version log with three columns: what changed, who approved it, and when. In a series that runs for a year, this document will save more time than any editing shortcut.

Organize assets before the first scene

Create folders for scripts, voice settings, brand assets, screenshots, music, and exports. Collect approved fonts, logos in light and dark variants, and a colour palette with named roles before you start building. Twenty minutes of asset gathering prevents the familiar mid-project hunt for the correct version of the logo.

Write Scripts That Survive Being Read Aloud

Structure that scales to any length

Use four beats and stretch them to fit the runtime:

Hook. The problem, the stakes, or a surprising fact. One or two sentences.
Promise. What the viewer will be able to do by the end. One sentence.
Process or proof. The body. Three to five ideas, each with supporting detail.
Action. The specific next step, phrased as a verb.

For a 60-second clip, each beat is a single sentence. For a ten-minute training module, each beat becomes a scene group with its own internal arc.

Rewrite for spoken rhythm

Read the draft aloud with a timer, then make these passes:

Replace semicolons and em dashes with full stops. Synthetic voices handle short sentences far more gracefully than nested clauses.
Split any sentence longer than roughly 20 words.
Convert passive constructions into active ones. Instead of saying that a report is submitted by a manager, say that the manager submits the report.
Write numbers and acronyms the way you want them pronounced on first use, then decide how they should read for the rest of the script.
Delete throat-clearing. An opening like an extended explanation of what the session will cover becomes a single line stating the three things the viewer will learn.

Build a pronunciation glossary

Brand names, product names, technical terms, and non-English words are the most common source of embarrassment in synthetic narration. Keep a side document listing the correct pronunciation of each term and apply the same settings to every project in the series. Consistency between episode one and episode seven is a large part of why a series feels professional.

Example: the same point, three ways

Weak version: a long conditional sentence explaining that in the event a session should time out due to inactivity, the user will be required to re-authenticate before continuing.

Better version: sessions time out after 30 minutes of inactivity, and users sign in again to continue.

Best version for narration: thirty minutes of inactivity ends the session, and signing in again takes about two seconds.

The third version is shorter, uses a concrete number, and gives the viewer an emotional anchor. It also gives the voice a natural place to pause.

Cast the Avatar and Design the Voice

Avatar selection is casting, not decoration

Viewers forgive imperfect rendering far more readily than a mismatch between presenter and message. A cheerful stock presenter delivering a serious data-breach notice undermines the content before a single word lands.

Stock avatar or custom avatar?

Consider these dimensions before deciding:

Setup effort. Stock avatars take minutes. A custom avatar requires recording sessions, review, and iteration measured in days or weeks.
Trust signal. A custom avatar of a named leader or instructor carries weight that a generic presenter cannot.
Reuse economics. Custom avatars make sense when the same person will appear in dozens of clips across months.
Maintenance. A custom avatar needs re-recording if the person's appearance changes materially.
Best fit. Stock avatars suit explainers, tutorials, and localized versions. Custom avatars suit recurring hosts, executive updates, and certification content.

A practical middle path: stock avatars for the bulk of internal content, one custom avatar where a specific person is genuinely part of the message.

Shooting a custom avatar well

Record in a quiet room with lighting on the face and torso that matches. Mismatched exposure reads as artificial.
Avoid fine patterns and stripes in clothing.
Speak at your normal presentation pace so the model learns realistic mouth shapes.
Keep gestures moderate. Hands crossing in front of the face are the most frequent source of visual artifacts.
Capture a range of expression rather than one neutral read, so the rendered avatar can carry the emotional beats your script requires.

Lock the frame and wardrobe

Choose a standard framing, such as a medium close-up with the eye line in the upper third, and reuse it. If half your videos are wide shots and half are tight crops, the series looks assembled from different productions. Do the same with wardrobe: two or three variants, rotated deliberately so the change reads as intentional.

Voice: synthetic, cloned, or recorded

Synthetic. Best for scale, multilingual output, and content where the narrator is not the message. Neural voices handle pacing well when sentences are short and punctuation is deliberate.
Cloned. Useful when one specific voice must stay consistent but recording time is scarce. Cloning requires clean source audio and a script that matches the speaker's natural phrasing.
Recorded. Still the strongest option for high-trust moments, including sensitive announcements, executive messages, and anything where listeners are judging sincerity rather than information.

Prosody and pacing controls

Punctuation is your main control surface. A comma produces a micro-pause, a full stop produces a beat, a paragraph break produces a breath. Beyond that:

Use shorter sentences at emotional peaks and longer ones during explanation.
Insert silence between major sections explicitly. Half a second before a new idea signals structure.
Do not stack emphasis. If every sentence is stressed, nothing stands out.
Pair pace with content: faster for lists and summaries, slower for warnings and instructions.

Music and sound design

Keep music at roughly 15 to 20 percent of perceived loudness beneath narration, instrumental only whenever speech is present. Fade the bed out three to five seconds before the final call to action so the ask lands in silence. Use transition sounds only when they carry meaning, such as a scene change in a technical walkthrough.

Build Scenes With Templates, Brand Kits, and Layout Rules

Create a brand kit once

Lock down and upload: two font families with defined sizes; a colour palette with named roles such as primary, secondary, accent, background, and warning; logo variants for light and dark backgrounds; a lower-third style; an end-card layout; and standard transition types with fixed durations. Applied through templates, this kit means a new video inherits your visual identity automatically instead of being styled by hand.

Use templates without looking templated

The fastest quality improvement available to any team is to break the template's rhythm on purpose: change the scene count, move one text block off the grid, swap a stock graphic for a real screenshot. Templates supply consistency. Small irregularities supply authorship.

Layout safety rules

Keep essential text 5 to 10 percent away from every edge, because exports crop differently across platforms.
Reserve a subtitle zone that never collides with the avatar's hands or a lower third.
One idea per frame. If a scene contains both a talking presenter and dense data, split it into two scenes.
Keep on-screen text short enough to read in the time the narration gives it. A useful rule: never display more text than you could say aloud in the same duration.

Scene modularity pays for itself

Keep scenes short and self-contained, ideally one sentence of narration each. Short scenes re-render quickly, localize selectively, and survive revision requests with minimal disruption. A scene that contains four ideas is a scene that will be rebuilt from scratch the first time one of those ideas changes.

The Three-Pass Editing Routine

Treat the first render as a rough cut, never as a deliverable.

Pass one: structure

Watch straight through without pausing and ask of each scene: if I removed this, what would be lost? Delete anything that fails the test. This is the highest-leverage edit available, and it is also the one most people skip because it feels like undoing work.

Pass two: timing and transitions

Confirm that every visual change lands on a spoken beat rather than an arbitrary moment. Trim lead-in silence, close awkward gaps, and keep transitions short, around half a second for corporate content. Watching at 1.5x speed makes timing problems obvious in a way normal playback does not.

Pass three: detail

Spell-check every on-screen text element.
Verify pronunciation of names, figures, and technical terms.
Check eye line and framing consistency between scenes.
Confirm that colours and fonts match the brand kit.
Normalize loudness across the whole video rather than scene by scene.

Accessibility as a standing requirement

Provide subtitles and a downloadable transcript. Captions raise completion rates on social platforms and are required in many accessibility contexts. Always break caption lines manually at clause boundaries, because automated line breaks routinely split a sentence across three stubby lines and destroy readability. Check both burned-in captions and any sidecar caption file.

Localization and Multi-Language Versions

Localization is where script discipline pays off. Short sentences turn translation into near-mechanical work: swap the voice, verify the avatar, re-render.

Practical guidance:

Localize rather than transliterate. Native reviewers should rewrite idioms instead of mapping them word for word.
Budget for length drift. Many target languages run 10 to 20 percent longer than English. Leave slack in scenes with tight timing.
Re-check embedded text. Text baked into images or screenshots cannot be translated automatically.
Decide about the presenter. For some audiences a locally recognized host increases trust. For others, a single consistent global host signals one company.
Keep a master matrix. Track which language versions exist, who approved them, and when they were last updated. Re-rendering is cheap. Re-approval is not.
Test with real viewers. Watch five seconds of a foreign-language version with a native speaker and ask whether the pacing feels natural. It is a fast and brutal check.

Export, Publish, and Repurpose

Export settings matter less than most people assume, but consistency matters a great deal. Standardize the presets and stop re-deciding them every time.

16:9 at 1080p or 1440p for learning platforms, web embeds, and presentations.
9:16 at 1080x1920 for social, with a frame that has been laid out again for vertical, never a centre crop of a wide video.
1:1 for feeds and email thumbnails.
Stills captured during export for thumbnails, decks, and help-centre images.
A separate audio track for audio-only versions and future reuse.

Build the repurposing plan into the project itself. One eight-minute module can produce a 60-second summary, three vertical clips, a transcript that doubles as a written article, and a set of stills. Planning that up front costs nothing. Retrofitting it later costs a day.

Mistakes, Troubleshooting, and Decision Criteria

The most expensive habits

Writing on camera. Composing the script inside the editor guarantees rewrites. Finish the script first.
One giant scene. Long scenes are hard to revise and impossible to localize selectively.
Default voice settings everywhere. The same voice at the same pace for a light onboarding clip and a serious compliance module flattens both.
Ignoring silence. Pauses are design elements, not dead air.
Skipping the transcript. You lose searchability, accessibility, and a free written draft of the same content.
No version log. Approvals get lost and people re-approve work that already shipped.
Rebuilding instead of re-rendering. When a few words change, edit the script and re-render that scene rather than starting the project over.

Fast diagnosis

Narration sounds rushed. The script is too long for the runtime. Cut around 20 percent of the words.
Avatar motion looks stiff. The writing is too formal. Shorten sentences and add pauses.
Lip sync drifts. One long continuous take with unusual phrasing. Split it into shorter segments.
Visuals feel inconsistent. Framing or wardrobe varied between scenes. Lock a standard frame and outfit set.
Captions are unreadable. Automatic line breaks. Re-break them at clause boundaries.
Localized audio overruns. No timing buffer. Add slack and re-check scene durations.
Video feels generic. The template was used untouched. Vary scene count, layout, and one visual element per project.

Decision criteria worth writing down

Before starting any project, answer four questions. Is this evergreen or dated? Does the message need a named human? Will this be localized within a year? Will it be repurposed into more than one format? The answers determine casting, scene count, script length, and export plan, and they take about two minutes to settle. Teams that answer these questions in writing before building almost never end up rebuilding.

FAQ

How long does a first avatar video take?

For a 60 to 90 second clip with an existing script, expect two to four hours including review. A ten-minute training module with custom graphics typically takes one to three days, most of which is scripting and layout rather than rendering.

Can I change the script after rendering?

Yes. Update the scene text and re-render that scene only. This is the strongest argument for keeping scenes short and modular from the beginning.

Is a synthetic voice acceptable for customer-facing content?

For explainers, tutorials, and support content where the information is the point, yes. For apologies, sensitive announcements, and high-stakes outreach, a real recording still performs better because listeners are evaluating sincerity.

How many avatars should one brand use?

One primary host plus one or two alternates for specific contexts is usually enough. Too many hosts dilute recognition. Too few create fatigue across a long series.

Do I need to shoot anything myself?

Only if you want a custom avatar or plan to cut in real footage. Screen recordings, product screenshots, and simple stock sequences cover most corporate needs without a camera.

How do I keep a series consistent over months?

Maintain a living style document covering palette, fonts, framing, voice settings, pronunciation glossary, scene naming convention, and a version log. Consistency is a documentation problem long before it becomes a creative one.

What is the fastest way to improve?

Watch your own video with the sound off, then listen with the screen off. If the visuals still communicate and the audio still makes sense, the structure is sound. Fix whichever half fails first.

How do I handle last-minute factual changes?

Isolate volatile facts in their own scenes and keep animated graphics free of numbers that change often. A policy update then means editing one line of script and re-rendering one scene instead of rebuilding the project.

The Real Skill Is Sequencing

The jump from beginner to pro in avatar video is not about discovering hidden features. It is about order: define the job, set constraints, write for the ear, cast deliberately, treat voice as design, build a reusable kit, review in passes, and plan localization and repurposing before the first render. Do that consistently and the tool disappears, which is exactly what good production should feel like.