{"html":"

Over a few days and six versions, an AI turned 451 holiday clips into a 15-minute family film and a 2-minute Instagram Reel. I never opened a video editing app. Here is how I went about it, what came out of it, and what it changes.

Twenty years of holiday films, always the same chore

This summer, we spent three weeks in Florida, from Miami to Orlando by way of Naples, Sarasota and Clearwater. When we got back, I had 451 clips on my phone: 1 h 56 of footage, mostly in 4K.

For 20 years, I have edited our holiday films every time, once or twice a year. I have used one app after another, but two things never changed: it took me a huge amount of time, and I struggled, because I am not an editor.

This time, I wanted to try something else. I already knew that Claude, Anthropic’s AI, could make videos: I had seen it produce motion design from an app’s source code. But I had never put it to the test on real content, shot with a camera.

So I handed it my footage, with a precise request: a punchy, vlog-style edit of 10 to 15 minutes, with music and cuts. It agreed, while telling me honestly what it would do less well than an editor.

The AI does the tedious work, I keep the decisions

Everything happened on my Mac, in a folder shared with Claude. The method comes down to eight steps.

  1. Set the stage. I dropped all the clips into a folder, without renaming them, along with the four music tracks I wanted. I added the trip itinerary, exported from Trip Control, our travel planning app, with every stop and its dates. The brief fits in ten lines.
  2. Take inventory. For each clip, the AI read the date, length, format and GPS position. Using each file’s creation date, it matched every clip to the right stop and city, spotted the videos received over WhatsApp and corrected for the Florida time difference.
  3. Watch the footage. The AI does not watch a video continuously. It extracts four frames per clip and assembles them into contact sheets, which it examines one by one. That is how it “saw” all 451 clips.
  4. Listen to what is said. A voice detector finds the spoken passages. Then a transcription model, Whisper, writes down word for word what everyone says, with a timestamp for each word. That is what makes it possible to cut at the end of a sentence and pick the best lines.
  5. Propose a storyboard. Before editing anything, the AI submitted a written plan: one chapter per stop, the highlights and the music for each part. This is the most important checkpoint.
  6. Edit. Each shot is trimmed, then synced to the tempo of the music. Vertical shots get a blurred background and the iPhone’s HDR colours are corrected. The music ducks automatically when someone speaks.
  7. Iterate. I watched each version and sent my feedback with timecodes, for example “at 4:24, my daughter is cut off”. The AI only re-renders the shots involved, so each new version arrives quickly.
  8. Finish. I approved a colour grade on a side-by-side comparison of frames, then the AI re-rendered the whole film from the 4K originals. Finally, it produced a vertical Reel with no family members in it, for Instagram.
\"Excerpt
An excerpt from the contact sheets: each clip is summarised in four frames, with its number, city, date, length and share of detected speech (“voix 99%” for clip 38).

Six versions and about fifty notes to go from decent to punchy

The final film is the sixth version. Between the first and the last, I sent about fifty notes, always in plain language and with timecodes. Three moments changed everything.

Reading the dialogue instead of guessing it

The first version already had a good rhythm, but the spoken passages felt off. Some cuts fell in the middle of a word, and the lines it kept were not always the right ones.

The reason was simple. The AI knew where someone was speaking, thanks to the voice detector, but not what they were saying. So it cut blindly, rounding each passage to the nearest beat of the music.

When I pointed this out, it diagnosed the problem and suggested installing Whisper. The automatic download was blocked on my machine, so it told me which four files to fetch myself, about 480 MB, and which folder to put them in.

In about twenty minutes, the 205 clips with speech were transcribed word for word. Three things changed:

That is when the edit went from decent to punchy.

Giving notes in plain language, as you would to an editor

I never touched a timeline. I watched the version, noted the timecode and described what I wanted. Two examples show how it works.

The basketball shot. At a funfair stand, my son tries to sink a basket. The first version cut before the ball came down. I simply wrote: “show the whole move, until we see the ball drop into the basket”. The AI extracted a series of frames from the clip to find the exact moment of the shot and of the ball going in, then re-cut the shot around that passage.

The University of Miami. I felt the campus sequence did not make enough of the footage. I told it what I wanted: more context, a caption with the place and date, and shots of the stands. It went through that day’s clips for shots of the stands, the shop and the park, worked them in around the spoken passage and extended the cut-off line to the end of the sentence.

The rest of the notes were along the same lines:

Each time, the AI found the right passage in the footage, checked it on screen and only re-rendered the shots involved. A new version arrived within minutes.

Choosing the colour grade on still frames

Once the edit was approved, I asked for a more cinematic look, with more vibrant colours. Rather than re-rendering 15 minutes of film for every attempt, the AI worked like a colourist:

  1. It picked six key frames from the film: a pool in full sun, a mural, a sunset, a night scene, a face outdoors and a group shot.
  2. It showed me three looks on a contact sheet, side by side with the original: a cinema teal & orange, a “vibrant summer” look and a “soft film” look.
  3. I asked for a blend of the first two, more restrained and with natural skin tones. It proposed a second sheet at two intensities.
  4. I chose the lighter one. It then re-rendered the whole film from the 4K originals, applying the grade before the titles so they stay pure white.
\"Colour
The second sheet: the original on the left, then the blend of the two looks at two intensities. I chose the light version, in the middle.

Two rounds on still frames, and the decision was made. The full render only happened once.

The result: a 14:52 film and a 2:06 Reel

Both videos come from the same 1 h 56 of footage.

The two edits delivered
Family filmInstagram Reel
Length14 min 522 min 06
Format16:9, Full HD9:16 vertical
Shots27895
Music4 tracks, chained on the beat1 track
Original soundKept on 6 min of spoken passagesNone
Versions before approval63

The film follows the trip in chronological order, in five chapters titled with their dates. The key places get a caption, such as Wynwood, the University of Miami, the Barbie house or Universal Studios. It ends with our end-of-trip interviews, then a recap where 85 photos drop onto one another to the beat of the music.

The Reel only shows scenery: sunsets, murals, cars, Disney Springs, Universal. I preferred not to expose my family on a public network.

I went from editor to director

I never moved a clip on a timeline. My job was to say what I wanted, then to watch and correct. The notes looked like the ones you would give a human editor: “group the breakfast together”, “let the shot pan all the way up”, “we don’t see some family members enough”.

The time goes somewhere else. Sorting 451 clips, transcribing the dialogue and syncing every cut to the tempo would have taken me several weekends. Here, my time went into watching and deciding. The renders ran while I did something else.

Transcription makes all the difference. As long as the AI picked spoken passages without understanding what was being said, it sometimes cut mid-word. As soon as it could read the dialogue, cuts landed at the end of sentences and the lines it picked became the right ones.

It is the same lesson as in my 75 days with a personal AI assistant: delegating the execution works all the better when you keep control of the decisions that matter.

The limits are real, and you need to know them

Five rules if you want to try

AI removes the chore, not the eye

This personal project illustrates exactly what I stand for with Levolia: AI does not replace the human eye, it removes the chore that kept us from getting started. It is also what I experienced when I built my first iPhone app without being a developer.

Those 451 clips would have stayed asleep on my phone. Today, they have become a film my family watches.

","dek":"After twenty years of editing our holiday films myself, I handed 1 h 56 of footage shot in Florida to Claude: the eight-step method, the six versions, and what you still have to do yourself.","quickTake":["From 451 clips, or 1 h 56 of footage, the AI delivered a 14:52 family film (278 shots) and a 2:06 Instagram Reel, without me touching a timeline.","The brief fit in ten lines: the clips, not renamed, four music tracks and the dated trip itinerary, which made it possible to match each clip to its stop.","The turning point was transcription: with Whisper, the 205 clips with speech were transcribed in about twenty minutes, and cuts stopped falling in the middle of a word.","The final film is the sixth version, after about fifty plain-language notes given with timecodes; the colour grade was chosen on sheets of still frames before a single full 4K render.","The AI judges from four frames per clip, sometimes misses a key moment and needs the machine to stay on: insist on a written storyboard and keep the final say on what matters."]}