avatar_spokesperson_production v1.0.0

---

name: avatar_spokesperson_production description: 'Produces avatar spokesperson videos people can trust: vendor-neutral workflow for deciding if a synthetic presenter fits the job, writing scripts at a pace avatars deliver well, getting written likeness and voice consent, keeping background and wardrobe consistent across a series, and placing disclosure where viewers actually see it. Use when planning, scripting, or reviewing an AI avatar, digital replica, cloned voice, or synthetic presenter video for marketing, training, onboarding, or social content.'

---

Avatar Spokesperson Production

Purpose

Make a synthetic presenter video that holds up after publishing: the right job for an avatar, a script written for how avatars speak, consent on file before anyone's face or voice is cloned, a look that stays consistent across episodes, and disclosure a viewer cannot miss.

This skill is vendor neutral. Avatar tools differ in pricing, render limits, languages, and avatar quality, and those details change often. Check the current docs of the tool in use at production time and record the tool and version in the production plan. What does not change is the order of work here: fit check, consent, script, consistency lock, disclosure, review.

Read references/consent-and-disclosure-checklist.md before any clone is created. Run bin/script-timer.py on every script draft.

Workflow

1. Decide if an avatar fits the job

• Use an avatar when the content is informational and repeatable: product walkthroughs, feature updates, onboarding steps, training modules, FAQ answers, internal updates, and the same message in several languages.

• Use an avatar when the presenter needs to be available on demand, the script changes often, or filming a person for every revision would stall the work.

• Use a human take instead when any of these are true:
• The video is a testimonial or review. A synthetic person cannot have tried a product, and presenting one as if it had is deceptive.
• The claim needs personal credibility: an apology, a safety notice, a health or financial explanation where the audience is judging the speaker, not just the facts.
• The moment is news, grief, crisis, or a first introduction where trust is the whole point.
• The script relies on real reaction, improvisation, or handling a live audience.

• If the job is mixed, split it. Let the avatar carry the stable explainer sections and film a person for the opening, the testimonial, or the closing commitment.

• Write the fit decision down in one line in the production plan, with the reason. If you cannot state the reason plainly, default to a human take.

2. Lock consent before creating any clone

• Get written consent from the person whose face, voice, or likeness will be cloned, before recording source footage or training a model. Verbal permission on a call is not consent on file.

• Cover these points in the consent record, at minimum:
• Who the person is, and who is allowed to create and use the clone.
• What the clone may say: topic limits, product limits, and any subjects it may never cover.
• Where it may appear: named platforms, paid ads, internal use, client work.
• How long the permission lasts, and how the person can revoke it.
• Whether the clone may be used after the person leaves the role or company.
• Pay or other consideration, if any, and who owns the source recordings and the trained model.

• Never clone a person who has not consented, including public figures, customers, employees recorded in meetings, and people in stock or archival footage. Tennessee's ELVIS Act, effective 1 July 2024, treats a simulated voice that is readily identifiable as that person's voice, and similar right of publicity rules exist in other states. The safe rule everywhere is the same: no consent, no clone.

• For a fully fictional avatar that resembles no real person, record that fact in the plan: source is a stock avatar or an original design, no real person's likeness was used. Keep the stock license or design record with the project files.

• Store consent records with the project, not in a chat thread. A video that cannot point to its consent record does not ship.

3. Write the script for an avatar, not for a person

Avatars deliver flat when the script is written like an essay. Write for the ear and for the render.

• Plan runtime from word count. Clear presentation speech runs about 125 to 150 words per minute. Plan avatar scripts at 130 words per minute to leave room for pauses the tool adds between sentences.
• 30 seconds: about 65 words
• 60 seconds: about 130 words
• 90 seconds: about 195 words
• 2 minutes: about 260 words
• 5 minutes: about 650 words

• Run bin/script-timer.py script.txt to get the word count, the estimated runtime at 130 wpm, and flags for long sentences and long words in a row. Fix flags before rendering, not after.

• Keep sentences short, mostly 8 to 18 words. One idea per sentence. Avatars handle a full stop better than a comma chain, and short sentences give the mouth shapes and pauses somewhere natural to land.

• Put the point in the first two sentences. Avatar videos lose viewers fast when the opening is a greeting and a self introduction. Open with the viewer's problem or the result, then say who is speaking.

• Write numbers and names the way they should be spoken. Write "forty nine dollars" if the tool misreads "$49", spell out acronyms on first use, and add pronunciation notes for product and person names in a comment block at the top of the script.

• Avoid lines an avatar cannot deliver honestly. No personal anecdotes for a fictional avatar, no "I tried this myself", no emotion the face cannot show. If a line needs real feeling, that section goes to a human take.

• Chunk long videos. For anything over about 90 seconds, write in named sections with a visual change between them: a new slide, a product shot, a caption card. A single unbroken avatar shot for five minutes reads as a lecture and hides edit points you will want later.

Worked example, product update script, 104 words, about 48 seconds at the planning rate:

"Your exports were failing when a file name had a space in it. That is fixed. Starting today, exports keep the original file name, spaces included, and the download link stays live for seven days instead of one. If you run scheduled exports, nothing changes for you. The new link window applies on the next run. One more thing. The exports page now shows why a job failed, in plain words, next to the retry button. No more guessing from an error code. Try it with your next report, and tell support if a link expires early. We read every report that comes in."

Note what it does: problem first, no greeting, short sentences, no claim the avatar personally used anything, one clear action at the end.

4. Lock the look and keep it consistent

Consistency is what separates a series from a pile of one off renders.

• Create a series lock sheet before the second video, and reuse it for every episode:
• Avatar identity: same avatar, same age and styling, no swapping mid series.
• Wardrobe: one outfit per role or series. Change outfits only when the series format itself changes, and note the change.
• Background: one background per series, or one small set of approved backgrounds. Keep the background plain enough that captions and product shots stay readable.
• Framing: same shot size every time, usually medium close up, head and shoulders, eyes near the upper third of the frame.
• Voice: same voice, same speed and pitch settings. Record the exact settings in the lock sheet, because a re-render with a default speed quietly changes the presenter.
• Captions: same font, size, and placement. Captions are part of the look, not an afterthought.

• Keep the avatar's claims matched to the visuals. If the script says "watch me open the settings page", show the settings page. Do not leave the avatar talking over a blank background while describing a screen the viewer never sees.

• Render a 10 second test with the lock sheet settings and watch it at phone size before rendering the full video. Check mouth sync on words with p, b, and m sounds, check hands if the avatar gestures, and check that captions do not cover the mouth.

5. Place disclosure where viewers see it

Disclosure is not a credit line. It has to arrive before or as the viewer forms an impression, and it has to survive the way people actually watch: muted, clipped, and reposted.

• Cover all four placements for a realistic avatar video:

1. In the video, early: on screen text in the first 5 seconds that says the presenter is AI generated, held long enough to read. For a cloned real person, name the arrangement plainly, for example "AI version of [Name], recorded with permission".
2. In the audio or captions when the endorsement or claim is spoken, for viewers who only hear it or only read it. The FTC guidance for video endorsements is to disclose in both audio and visual form because some viewers watch without sound and others miss on screen text.
3. In the description or caption, in the first lines, not after a "read more" break and not buried in hashtags.
4. In the platform's own tool at upload: YouTube Studio AI use set to Yes for realistic synthetic content, TikTok's AI generated content toggle, or the equivalent on the platform in use. A platform tool does not replace the in video disclosure, and the in video disclosure does not replace the platform tool. Use both.

• Add a separate ad disclosure when the video promotes a product or service for pay or other material connection. "AI presenter" and "ad" are two different facts and both need saying, in simple language. The FTC revised Endorsement Guides from June 2023 cover virtual influencers and require material connections to be clear and conspicuous, meaning difficult to miss and easily understood.

• For audiences in the EU, Article 50 of the EU AI Act requires deployers of deep fake image, audio, or video content to disclose that it was artificially generated or manipulated, clearly and at first exposure. The in video first seconds placement above is built to meet that timing. These transparency duties apply from 2 August 2026.

• Never use disclosure to launder a claim an avatar should not make. Labeling a fake testimonial does not make it a real one. Cut the claim instead.

6. Review the render like a skeptic

• Watch the full render once without pausing, at phone size, with sound on. Note every moment you stopped believing it or stopped following it.

• Check these specific failure points:
• Mouth sync drifts after a cut, a long word, or a number.
• Hands repeat the same gesture on a loop, or fingers deform in a gesture shot.
• Eye direction shifts, blinking looks timed, or the head moves on a pattern.
• A pause lands mid phrase, or emphasis lands on the wrong word and changes the meaning.
• The background, wardrobe, or voice differs from the lock sheet.
• Captions misspell a name or number that the voice says correctly, or the reverse.

• Fix by editing the script first. Most sync and emphasis problems are sentence problems. Shorten the sentence, split it, or spell the word phonetically, then re-render only that section if the tool allows section renders.

• Get a second person to watch cold and ask them two questions: what was the video asking you to do, and when did you learn the presenter was AI generated. If they cannot answer the second one, the disclosure failed, no matter where it was placed.

Output Contract

Deliver all of the following for each video or series:

• Production plan: job, fit decision with reason, tool and version, avatar source, and target runtime.
• Consent record on file, or the fictional avatar note, with its storage location named in the plan.
• Final script with word count and estimated runtime from bin/script-timer.py, plus pronunciation notes.
• Series lock sheet: avatar, wardrobe, background, framing, voice settings, caption style.
• Disclosure record: the exact in video wording, its timestamp, the description wording, and the platform tool setting used at upload.
• The rendered video with captions, plus the review notes from the skeptic pass and what was fixed.

Operating Rules

1. No consent, no clone. Written consent comes before source recording or model training, every time, with no exception for internal only videos.
2. Avatars explain, people testify. Never script a synthetic presenter to claim personal experience, personal use, or personal feeling it does not have.
3. Disclosure goes in the video, in the description, and in the platform tool. One of the three is not enough for a realistic avatar.
4. Plan at 130 words per minute and keep sentences short. If the timer flags it, fix the script before paying for a render.
5. One series, one look. Any change to avatar, wardrobe, background, or voice settings goes on the lock sheet with a date and a reason.
6. When the job needs trust more than it needs scale, film a person. Say so in the plan instead of forcing the avatar to carry a message it weakens.
7. Verify tool specific limits, languages, and pricing in the tool's current docs at production time. Do not carry vendor numbers from memory into a plan or a quote.
8. Keep every factual claim in the script to the same standard as any other video: sourced, checkable, and cut if it cannot be verified.