Explainer Video

Explainer Video

Turns a subject under discussion into a narrated, captioned MP4 that explains it. The output is a finished video in output/, not a deck and not a script.

When NOT to Use

• Slides for live presenting → pptx. A deck is not a video.
• A single image, diagram or infographic → image-operations / pptx.
• Audio only (podcast, briefing) → call PodcastGenerate directly.
• A written explainer (SOP, guide, one-pager) → docx.
• Marketing film, live action, real people's likenesses, licensed footage or music — not possible here. Say so plainly and offer the b-roll style this skill can do.

Workflow

1. Frame the subject (one short round)

Establish, from the conversation first and only then by asking: subject, audience (newcomer / practitioner / exec), target length (default ~2 minutes ≈ 6-8 beats), and the angle ("when to use each option", "how it works", "what changed"). Use AskUserQuestion once — never a chain of questions.

Always ask the visual style in that same card: dark (navy slides, colour-coded sections, photographic b-roll) or light (white slides, brand accents, sparing photography).

Ask whether they have screenshots to include if the subject is a product, tool or UI walkthrough.

2. Gather (only what the video actually needs)

• Research when it matters — external facts, product capabilities, statistics, "what's new": use web_search / web_fetch, or the deep-research-agent when claims need citing. Skip research entirely when the subject is already covered by this conversation or the user's own content. Never research a topic the user has already explained.
• Internal subjects — pull from SearchM365, ReadFileContent, meeting transcripts. Ground every internal claim in what those return.
• Screenshots — Glob input/**/* and check any <attached_files> block. Screenshots are used as supplied: framed, optionally box-highlighted, never regenerated or redrawn, so the product UI stays truthful. Note any UI text you rely on in narration so it matches the pixels.

3. Write the script — and STOP for approval (mandatory gate)

Build a storyboard: one beat per scene, each with its narration sentence(s). Rendering takes ~10-15 minutes, so the script is approved before any pixels are generated.

Show the user, inline in chat: the beat list (title + one-line narration each), the estimated runtime (≈ 150 spoken words per minute), and the style chosen. Ask for a go-ahead or edits. Do not generate images, audio or video until they approve.

Write the approved storyboard to working/storyboard.json:

{
  "style": "dark",
  "beats": [
    {"kind": "title", "photo": "working/broll-wide.png", "title": "…", "sub": "…",
     "foot": "…", "narration": "…"},
    {"kind": "section", "accent": "blue", "tag": "1 / 4", "photo": "working/broll-1.png",
     "title": "…", "kicker": "…", "bullets": ["…", "…", "…"], "example": "\u201c…\u201d",
     "narration": "…"},
    {"kind": "screenshot", "accent": "green", "image": "input/shot1.png", "title": "…",
     "caption": "…", "callouts": [{"x": 40, "y": 120, "w": 400, "h": 90}], "narration": "…"},
    {"kind": "table", "photo": "working/broll-team.png", "title": "…",
     "rows": [{"left": "…", "right": "…", "accent": "blue"}], "narration": "…"},
    {"kind": "closing", "photo": "working/broll-team.png", "title": "…", "sub": "…",
     "foot": "…", "narration": "…"}
  ]
}

Accents: blue, green, purple, orange, teal, red — one per section, reused by its cards.

4. Generate the b-roll

One photo per beat that has no screenshot, via ImageGenerate (quality="medium", orientation="landscape", size="large", destination="working"). Issue the calls in parallel — each takes ~2 minutes. Prompt pattern that works:

Cinematic corporate b-roll photograph: , modern office, tones, soft daylight, shallow depth of field, professional stock-photography look, screen content abstract and not legible, no text or logos.

Vary the scene, the people and the colour per section so the video does not repeat itself. Never ask for real individuals, real brand logos, or legible on-screen text (the model misspells it).

5. Record the narration

PodcastGenerate with format="single_host" and one turn per beat, in beat order — the renderer detects the pauses between turns to cut the scenes. Keep each turn conversational and self-contained; write numbers and symbols as words. It returns the MP3 path and duration.

6. Compose and render

python scripts/compose_cards.py working/storyboard.json --out working/cards
python scripts/render_video.py working/storyboard.json --cards working/cards \
  --audio output/<narration>.mp3 --out working/<name>.mp4

compose_cards.py builds 1920x1080 cards (section beats get a b-roll interlude plus a detail card). render_video.py aligns scene changes to the narration pauses, applies slow Ken Burns motion to photo scenes, crossfades, generates captions from the narration and burns them in. Rendering takes 10-15 minutes — run it with a long initial_wait and wait for it, never abandon it mid-flight.

Publish the finished file:

CopyArtifact(surface="output", source="working/<name>.mp4", destination="<name>.mp4")

Then Glob output/**/* to confirm it landed before telling the user it is ready.

Output Format

Deliver to output/: the MP4 (1080p, H.264 + AAC, burned-in captions), plus the narration MP3 and script text that PodcastGenerate already writes there. In chat, state the exact filename, the true runtime, and what each section covers — then offer the likely next edits (different b-roll, longer script, add screenshots, a light-style version).

Guardrails

• The script gate is not optional. Never generate images, audio or video before the user approves the beat list.
• On-screen claims must be true of the video itself. If a card says "a two-minute guide", the render must be about two minutes. Check the reported duration against every claim before publishing, and re-cut the card rather than shipping a mismatch.
• Never invent facts. Statistics, quotes, dates, product capabilities and customer names come from research or the user's content, or they do not appear. Where something is genuinely unknown, leave a visible [placeholder] on the card and say so — never fill it with a plausible number.
• Screenshots are evidence, not art. Show them as supplied. Never regenerate, retouch or "clean up" a product UI, and never fabricate a screenshot of a UI the user did not provide.
• People in b-roll are synthetic and generic. Never generate a real person's likeness, a real brand's logo, or copyrighted characters — offer an original equivalent instead.
• Reproduce nothing copyrighted in narration: no song lyrics, no verbatim article passages, no licensed music. Narration is original writing; quote a researched source in under 15 words with attribution, or paraphrase.
• Say what failed. If narration, image generation or the render fails after one retry, tell the user which part is missing rather than shipping a silent or half-built video.