Go from one idea to a finished 10-second vertical video with Claude as the director: story, shot list, a fresh-eyes review at every stage, voice, music, and a final check. The rule throughout: fix it in words before you pay for pictures, and in pictures before you pay for video. Inspired by a widely shared X post on AI video, rewritten with the steps and mistakes from how we make our own Reels.
I make a product called What it does in one line. Write 5 opening lines for a 10-second vertical video. Each must name a specific pain my user feels and include a real number they'd recognise. No product name in the first line. Then pick the best one and say why.
From our own Instagram: the 3 posts that opened with a pain and a number got 131 to 139 views each; posts that opened with a feature name got 60 to 70 (posts from late 2026-10 to 2026-11-07, compared by views per post). A very small sample, so treat it as a hint.
Turn this opening into a shot list for a 10-second vertical video. A table with: seconds, what's on screen, big on-screen text, voice line. 3 to 4 shots. The opening line is on screen in the first frame. End within 1 second of the last line.
Keep it to 3 or 4 shots. Every extra shot is another place for things to go wrong.
Open a new chat. Paste only the shot list, nothing else.
You are seeing this shot list for the first time. You are a viewer, not the writer. List everything that would confuse you or look wrong: objects that change or appear from nowhere, numbers you couldn't check on screen, text that's on screen too short to read, lines that don't match the picture. Don't fix anything, just list problems, most serious first.
Tried on Claude, 2026-11, on a 3-shot list we wrote: it caught a cup that changed colour between shots, a number the picture didn't show, and a gap in the timing. All fixed in words, at no cost.
Ask Claude for one page: colours (5 at most, with codes), light, camera height, what the main character or object looks like, and what must never change between shots. Every picture prompt after this starts by pasting the style sheet.
For each shot, ask Claude to write the prompt in the same seven parts:
1. Who or what is in it
2. What they're doing
3. How the shot starts
4. How it ends
5. Camera: where it is and how it moves
6. Light and style (from the style sheet)
7. What must stay the same as the last shot
Say what you want, not what you don't: picture models often ignore "no" and "don't".
In a fresh chat, give the reviewer the style sheet and the pictures, and ask it to compare them one pair at a time: same character, same object sizes, same colours, nothing new appearing. Only pictures that pass go on to video.
Music sits about 14 dB under the voice, and we set the whole thing to -16 LUFS. With ffmpeg:
ffmpeg -f concat -safe 0 -i shots.txt -c copy picture.mp4 ffmpeg -i picture.mp4 -i voice.wav -i music.wav -filter_complex "[2:a]volume=-14dB[m];[1:a][m]amix=inputs=2:normalize=0,loudnorm=I=-16:TP=-2[a]" -map 0:v -map "[a]" -c:v copy -c:a aac -shortest final.mp4
shots.txt lists your clips, one per line, as: file 'shot1.mp4'
`-c copy` in the first line only works when every clip has the same codec and size. If they differ it fails with an error; remove `-c copy` so ffmpeg re-encodes.
Tried in 2026-11 on three test clips: 9.0 seconds, measured at -16.0 LUFS.
Watch it twice in a fresh chat or with a friend:
- Sound off: can you read every line? Is the opening clear in the first second?
- Sound on: does any line sound like a question? Is the music muddy or harsh?
- Does it end within a second of the last line?
| What went wrong | Fix |
|---|---|
| The voice said "Cloud" instead of "Claude" | Spell it "Clawd" in the voice script, then check with speech-to-text |
| Long narration got muddy and hissy towards the end | Generate one line at a time |
| A statement sounded like a question | Ask for a falling end on every line, and listen for it |
| The ending dragged on | Stop within a second of the last line |
| The music sounded like an old radio | The high end had been cut; keep the highs |
From our own production notes, 2026-10.
| What goes wrong | Write it this way instead |
|---|---|
| An object changes colour or size between shots | Name it the same way in every prompt, with its colour and size |
| Something appears from nowhere | Say where it is at the start of the shot |
| Several people do the same move at once | Give each person their own action |
| "No text" gets ignored | Say what should be there instead |
Summarised in our own words; the idea of keeping a table like this comes from Eian's post.
Pick the pain your users mention most, run steps 1 to 3 only, and stop. Read the review. Most of the value is already there before any picture exists.
Opening lines, shot list, style sheet, review notes and the final file, as cards on one put·here canvas. The next video starts from a copy.
Pick AirDrop to send it to your Mac. No AirDrop? Email it to yourself or save it to Notes.
The phone app can’t import yet. Open this link on a Mac to put it in put·here.