Seven YouTube Shorts in 46 minutes: how an AI drove DaVinci Resolve for me

This morning I visited Kellers Ranch in Weiterstadt, an animal sanctuary with around 300 residents, and filmed with my Google Pixel 7 in “Cinematic Blur” mode. That is the mode where the phone artificially blurs the background, like a big camera would. I came home with 20 clips: camels, ring-tailed lemurs, a caracal, donkeys, alpacas, dogs and a few animals I had to look up first.

Normally that means hours in front of the editing software in the evening. This time I wanted to try something different.

Why?

I wanted to find out whether an AI can now really work with my programs by itself. Not just give tips (“click Edit, then …”), but actually cut, build titles, fetch music and spit out finished videos at the end. In DaVinci Resolve, no less, which is a real and fairly complex professional editor.

The brief was roughly: “Turn these clips into at least six YouTube Shorts that could go viral.”

So what is an MCP server?

This is the part that sounds like magic to non-nerds, but it is actually simple.

On its own, an AI like Claude is just a very clever conversation partner. It can write and think, but it has no hands. It cannot click anything in DaVinci Resolve.

An MCP server is something like a remote control with an interpreter. The maker of a program (or someone from the community) builds a small add-on that can give the program commands: “create a new timeline”, “put this clip at second 4”, “start the export”. The AI gets a list of these commands and can then use them itself. MCP stands for Model Context Protocol, which is simply the shared standard so any AI can talk to any program this way.

For that to work, the AI has to run on my PC rather than somewhere in the cloud, because DaVinci Resolve runs on my machine too. That is why I used Claude Code this time, which works directly on my computer. Three of these “remote controls” were involved:

• DaVinci Resolve for editing, timelines and export
• EffectCraft (from the ArtCraft suite) for the animated text overlays, although there was a catch, more on that in a moment
• Chrome, so the AI could pick music in the YouTube Audio Library

The AI set up the MCP connection to EffectCraft and PhotoCraft by itself, by the way. The catch: Claude Code can only use new connections after the session is restarted. Instead of waiting, the AI simply drove EffectCraft through its command-line tool, like typing commands instead of using the remote control. It worked just as well. It ended up not needing PhotoCraft at all.

The result: seven Shorts

Before the details, here is what came out the other end. All seven are vertical 4K, between 21 and 37 seconds long, and meant for my private channel. The on-screen text is in German.

How did it go?

For the first round I sent the AI just four messages. From the first message to the last finished export took 46 minutes, and that includes the time I spent looking at the concepts. A short correction round came later, more on that below.

1. Look first, learn second

First the AI went through the 20 clips. Since it cannot “play” a video the way we do, it pulled stills from every clip at three-second intervals and looked at those. That way it knew where the camel was, where the lemur was and where someone was scratching a dog.

Along the way it read up on what the Pixel's Cinematic Blur actually does. Result: the blur is baked into the footage while filming and cannot be changed afterwards. And the clips have no sound. I knew that, but the AI worked it out on its own.

2. Seven concepts straight in Resolve

Next it created seven separate timelines in DaVinci Resolve, one per Short, and cut the matching clips into them. No effects and no music, as I had asked, so I could judge the ideas first.

The nice thing: this is no special “AI format”, just a normal Resolve project. I can jump in at any time and keep working myself.

What I liked most is that the AI put a marker with stage directions on every cut: which text appears, which music mood fits, where the “hook” is, the moment meant to stop people from swiping on.

Then it asked me six questions: what kind of music, whether the ranch's website should appear in the video, whether there was original sound, whether I knew stories about individual animals, whether a tattooed forearm visible in the shot was okay, and whether to export in 4K. I answered briefly and approved all seven concepts.

3. Animated titles from EffectCraft

This is where it got exciting. For the text overlays the AI used EffectCraft. It built one test title first, checked the result and then wrote a small program that generated all 46 titles in one go: in 4K with a transparent background, so they sit on top of the video. Afterwards it placed every title in Resolve exactly on the matching cut.

At my request, “for Kellers Ranch” became “for my private channel” along the way: logo and website were dropped from the videos, and Short 05 now ends with a question instead of a call for donations.

4. Music from the YouTube Audio Library

I only wanted music from the YouTube Audio Library, so there would be no copyright trouble later. That meant the AI had to go into YouTube Studio, in the browser. At first that failed because the Claude extension in my Chrome was not properly connected. The second attempt worked: the AI filtered by mood, chose only tracks that need no attribution and downloaded seven of them.

It then trimmed the music to length, levelled the volume, added a gentle fade-out at the end and laid a track under every timeline.

5. Rendering

Finally it added all seven Shorts to Resolve's render queue and exported them: 2160 × 3840 pixels, 30 frames per second, H.265. Each Short took only a little over ten seconds.

6. One correction

When I watched the Shorts, I noticed that the lemur mum in Short 04 is not carrying a baby at all. It is simply a very curious lemur. So I sent the AI a single sentence. Three minutes later the Short was redone: four new titles, more playful music instead of the sad harp, and a new name. The edit itself stayed as it was.

I did not take the screenshots above myself, by the way. For this post Claude re-created the scenes in the real Resolve project and captured them.

What the AI cannot do (yet)

To be honest, and the AI was honest about this too:

• It cannot hear the music. The tracks were picked by title, genre and mood, not by ear. It still fitted surprisingly well.
• It never saw the videos “in motion”. Everything was checked with stills. That is why the cuts are not yet on the beat, and why in Short 07 the closing question sits over a path instead of the dog, because the camera pans away at that moment.
• Stills can be deceiving. In one lemur the AI saw a mum carrying a baby on her belly. There was no baby, just a very curious lemur (see step 6).
• Species identified by appearance. It was unsure about “Zwergzebu” (dwarf zebu) and “Löffler” (spoonbill) and asked me to double-check.
• It did not work entirely without me. I had to connect the browser and confirm a few save dialogs during the music download.

For a first pass it is still impressive. Polishing is polishing, but the hours of sorting, trimming and fiddling with titles were completely taken off my hands.

How many agents were involved?

For bigger tasks Claude Code can send out helper agents, additional AI instances that work on parts of the job in parallel. That was not needed here: a single agent did everything on its own, with no helpers. It made 136 tool calls, broken down like this:

The same model was used throughout: Claude Fable 5.2.

What did it cost?

Because I used Claude Code this time, I can read out the usage exactly. A small tool called ccusage analyses the Claude Code logs on my PC.

A quick explanation: AI models count in tokens. A token is roughly a word fragment, about three quarters of a word. Images count too, and quite a lot. Every time the AI takes a step, it rereads the entire conversation so far, including every still it has looked at. To keep that affordable there is a cache: anything already read once costs only a fraction when it is read again.

That is just over 17 million tokens for both rounds together, 98 percent of them from the cache. Had I paid through the regular Claude API, it would have cost 9.11 US dollars according to ccusage, or about 8.14 euros (ECB rate of 9 November 2026: 1 euro = 1.1193 dollars).

I did not actually pay anything extra, because I use Claude on the Pro plan for 21.42 euros a month. The correction of Short 04 accounted for only about 80 cents of that. On paper, this one evening “used up” a good third of the monthly fee.

Model used: Claude Fable 5.2. All the work was done by Fable 5.2, currently Anthropic's most capable generally available model. You can tell mostly from how independently it switched between programs and how precisely it said what it had checked and what it had not.

About Kellers Ranch

The stars of these Shorts live at the Tierhilfeverein Kellerranch e.V. in Weiterstadt, near Darmstadt. On around 10,000 square metres it is home to more than 300 animals in need: found animals, surrendered animals and animals rescued from poor conditions. Some are rehomed, others get to spend the rest of their lives at the ranch. There is also a parrot house, a lemur enclosure and, yes, camels, rheas and a caracal, right in the middle of Hesse. The association is a member of the German Animal Welfare Federation (Deutscher Tierschutzbund).

The ranch is open to visitors daily from 1 pm to 5 pm, and dogs on a lead are welcome. Address: Im Wasserlauf 3, 64331 Weiterstadt, Germany.

How you can help

A place like this runs on donations, sponsorships, memberships and volunteers. If you would like to give something:

• Donation account: Tierhilfeverein Kellerranch e.V., Frankfurter Volksbank Rhein/Main eG, IBAN DE63 5019 0000 0000 4589 70, BIC FFVBDEFF
• Donation receipts: the association issues them collectively in the first quarter of the following year.
• Other ways such as PayPal, sponsoring an animal or helping out are listed on the association's website under “Unterstützen”.

All the details, upcoming events and animals looking for a home are at kellers-ranch.de (in German). Please double-check the bank details there before transferring; the official site is always the most reliable source.

And if you are nearby: go. The nosy lemur is even better in person, and it will definitely come over.

Music in the Shorts: Morning Mandolin (Chris Haugen), Tiptoe Out the Back (Dan Lebowitz), Whistling Down the Road (Silent Partner), Touching Moment (Wayne Jones), Bike Rides (The Green Orbs), Dog Park (Silent Partner), all from the YouTube Audio Library.