Vox-style explainer, built in code
You make a 30-60 second documentary explainer as one HTML file, then export it to MP4. Every frame is code. No video generator, no stock footage.
Before you start: read brand.md and MOTION.md if they exist. If not, run brand-intake first, or ask for hex codes, a font and a logo file. Never call a result on-brand without them.
Step 1: Ask for four things, then wait
- The question. One "why" question a stranger half understands. Example: "Why does every brand logo look the same now?"
- The facts. Names, dates and numbers that answer it, each with a source link. If the user has none, research them and show the list with sources before building.
- The colours. Three hex codes plus one accent. Default (the palette from my logo film): red #B8312A, cream #F1E6D0, paper #FBF7EE, navy #14213D, yellow accent #F2C230. Label defaults as defaults.
- The size. 1920 x 1080 (YouTube), 1080 x 1350 (LinkedIn feed) or 1080 x 1920 (Story or Reel).
Do not start building until every fact on screen has a source.
Review rules (from the rounds that got rejected)
- Sound is not optional. A silent explainer feels unfinished. Write a score in code: a low bed, a hit on each new shot, a lift on the last line.
- Cards and images, not lines. A single bar or line moving across the screen reads as static. Every shot uses cards, real images or objects that move in depth.
- Logos fill their card and come from the real files. A logo floating small in a big box looks wrong.
- Keep grids symmetrical. Equal card sizes, equal gaps. A lopsided layout reads as a mistake.
- The voice sits on top. If you can't hear the narration over the effects, drop the score until you can. A voice track (your own, or a text-to-speech one) carries a "why" film.
- The opening line has to promise what's coming. A small title that undersells the film loses people before the good bit. Say the question big, then show the first surprise within 5 seconds.
- Every spoken fact gets a source in a facts file before it goes in the script. Say "looks like", never "is", when the research only shows a similarity.
- Stunning or it's not done. If the first 3 seconds would not stop your own thumb, rebuild them before building the rest.
Step 2: Write the script, one line per shot
- 8-14 shots. One sentence each. Shot 1 is the question itself.
- Every shot answers something the line before it said. If you cannot say what a shot proves, cut it.
- The last shot is the takeaway, in five words or fewer.
Show the script as a numbered list and wait for a yes.
Step 3: Pick a shot type for each line
| Shot | Use it for |
|---|---|
| Kinetic statement | One sentence you want remembered. Three short lines of bold condensed type on a torn paper strip. |
| Paper cards | Showing several things at once (logos, products, quotes). Torn-edge cards drop in with small shadows. |
| Before / after | A change over time. Same card, date chip flips. |
| Label call-out | Naming one thing inside a picture. A marker circle draws itself, then a label. |
| Counter | A number that grows. Built in code, never guessed. |
| Grid | "They all look the same." Many identical tiles, then one breaks the pattern. |
Step 4: Build it
Paste this build brief into Claude Code with the script and shot list underneath:
Build one self-contained HTML file called index.html. No external fonts, images or libraries unless I attached them. Nothing loaded from the internet.
Look: editorial documentary explainer. Palette strictly [RED], [CREAM], [NAVY] with a [YELLOW] accent. Torn paper cards with small drop shadows. Halftone dots, paper grain and fine film grain over the whole frame. Colour fringing at the frame EDGES ONLY, and a soft vignette. Bold condensed sans-serif type, fully formed as blocks.
Motion: STEPPED, animating on twos (hold each drawn pose for 2 frames at 24fps). Never smooth corporate easing. NO typewriter effect. No gradients on text, no glow, no bouncing.
Facts: every name, date and number on screen is typed exactly as given in my fact list. Add nothing I did not give you.
Timing: follow my shot list. Each shot holds long enough to read aloud.
Build rules: work out every frame from the time alone. Add window.seek(seconds) that draws that exact moment. The last frame can differ from the first (this is a film, not a loop).
After the build: capture one frame from the middle of every shot. Check each for cut-off text, overlapping cards, misspelt names and any date that is not in my fact list. Fix what is real and keep the first version.
Step 5: Check it like an editor
Before export, go through every frame capture and confirm:
- Every date and name matches the fact list, character for character.
- No card covers another card's text.
- Nothing is smooth that should be stepped.
- The first 3 seconds already show the question (people decide in 3 seconds).
Step 6: Export
Easiest route: HyperFrames, free and open source from HeyGen (npx skills add heygen-com/hyperframes, then npx hyperframes render). Any tool that steps through window.seek and saves frames as MP4 also works. Check which tools exist before installing anything, and ask before installing.
What this skill will not do
- Invent a fact, a date or a quote.
- Recreate a real brand's logo from memory. Use the user's files or plain type set in the brand's name.
- Claim a trend is caused by one thing. Say "one reason", not "the reason".