Turn a help article into a 45-second avatar script and render it

Take one task from the help article, write the steps as spoken sentences of about 100 words, and render that as a 45-second Sume avatar clip. At Standard, 45 seconds is 45 x $0.184 = $8.28. Do not read the article aloud: a clip should do one job, and the article stays the reference for everything else.

What to keep and what to cut

Articles are written for scanning: headings, tables, warnings, and links. Speech is linear. Keep the goal, the three or four steps in order, and one common mistake. Cut anything with a long path name, a code block, or a table, and put a sentence in the clip that tells people where the article is.

Sume estimates the video duration from the script, and a script that estimates outside 4-60 seconds is rejected. A rule of thumb that is easy to check: if the script is more than about 150 words, split it into two clips.

Render with a preview

Submit the script to the preview route first. You get a first-frame still, and generate-video renders it on approval. Preview stills are tier-independent, so you can approve at one tier and choose the final tier at render time.

import json

article_steps = [
    "To reset your password, open Settings and choose Security.",
    "Select Reset password, then check your email for the link.",
    "The link works once. If it expired, request a new one.",
    "The full guide is linked under this video.",
]
body = {
    "avatar_handle": "help_host",
    "aspect_ratio": "16:9",
    "quality": "standard",
    "script": " ".join(article_steps),
}
print(json.dumps(body, indent=2))

Keep the clip and the article in sync

• Store the article URL, last-edited date, and avatar_video_id in one row.
• When the article changes, flag the clip for review; do not auto-render.
• Put the transcript on the page too. Sume returns metadata.transcript_text after a render, which helps with search and accessibility.
• Add captions for people who watch without sound.

A worked example

Say the article is a ten-step guide to exporting a report. The clip should not walk all ten steps. Pick the step people get wrong, usually the one with a setting that is hard to find, and speak it in two or three sentences: where to click, what to choose, and what they should see afterwards. Then end with a pointer to the full guide. That clip is about 40 words and lands near 15-20 seconds, so you can pay about $3 to $5 for a very targeted answer, instead of $8 for a clip that tries to do everything.

Then look at your support tickets for the article. If one step produces most of the tickets, make a clip for that step first. The tickets give you an evidence-based priority list, and a before-and-after count tells you if the clip helped, which is a better test than views.

When not to

Skip the clip for pages that change weekly, for security or legal text, and for anything that needs exact commands the reader must copy. For those, the page is the product.