I Couldn't Learn After Effects, So I Built My Own. It Taught Me Product.

The problem

Here's a confession: I love motion graphics, and I could never learn After Effects.

I wanted to make catchy short videos. A stat that slams onto the screen. A lyric video. A quick promo for my own app. Every time, I'd open After Effects, stare at a wall of panels, layers and graph editors, watch a tutorial or three, and close it with nothing made.

I had the ideas. I just couldn't turn them into motion. So instead of learning After Effects, I built my own motion graphics studio, where an AI agent does the keyframes and I do the taste. What I didn't expect was that building it would teach me more about product than about animation.

Not another "type a prompt, get a video" app

The obvious move was a text-to-video model. I didn't want one. Generated video is a slot machine: you pull the lever, get something close, and then you can't say "move that word up 20 pixels." You just regenerate and hope.

So instead, the agent operates a real animation engine. Every video is one scene file with layers, keyframes and a camera. The agent edits that file using the exact same actions I use in the UI. That means everything it makes is editable, frame accurate and consistent from one clip to the next. If one word is off, I fix one word, and nothing else moves.

The agent isn't the renderer. It's the crew.

The deal: I judge, it executes

I split the work by what each side is actually good at. I'm good at deciding what a video is about, pointing at what matters, picking between options and reacting to a preview. I'm bad at (and bored by) easing curves, layout math and timing forty layers by hand. The agent is the opposite.

So I talk to it in a chat panel, it builds the scene live, and every turn it takes is one undo step. There's also a Direct mode: I drag the camera around roughly, like waving my phone, and the agent cleans it up into a smooth, precise move.

One rule I'm proud of: user edits are sacred. If I touch something, the agent can't overwrite it unless I explicitly say so. That's enforced in the code, not just asked nicely in a prompt.

Teaching taste with rules

"Make it catchy" isn't an instruction anyone can follow, human or AI. So I wrote down what catchy actually means: hook the viewer in the first second, overshoot and settle, stagger things, keep something always moving, mix fast and slow, and leave text up long enough to read.

Then I gave the agent a way to check itself. After it builds something, it renders a contact sheet of its own frames and asks four questions:

Is the first frame dead?
Is anything sitting still too long?
Can every line be read in time?
Is too much happening at once?

Good motion doesn't come from hoping the model has taste. It comes from rules plus a mirror.

Then it got a little out of hand

Once one agent could build a video, I wondered what would happen if I let a few of them loose. Now I have a newsroom: six agents, two each for YouTube Shorts, Instagram Reels and TikTok.

Each one picks its own niche, researches facts on the web, pitches an idea with sources, builds the video, checks its own frames and drops the finished clip into a Slack-style channel for me to review. They've come up with titles like "$7.25 hasn't moved in 17 years" and "A Day on Mercury Lasts Two of Its Years."

So far they've made 25 videos at about $1.70 of model usage each, all running on my Claude Max plan. A few things stay human on purpose: every fact needs a source, the agents only use material we own, and I'm the one who hits post.

The product calls behind it

The code was the easy part to delegate. The hard part was deciding what to build, what to skip and what each choice would cost. Here are the calls that shaped the studio most.

Bring it all into one app

My first plan was modest. I'd learn just the slice of After Effects I needed, and let the studio make graphics while DaVinci Resolve handled the rest. But every hop between apps cost time, so one by one the pieces moved in: MP4 export, music, footage, data charts. Today the studio assembles almost the whole video, audio included. Keeping everything in one place simply made more sense than stitching tools together.

That's when it clicked. After Effects probably went through the exact same thing. Every feature makes sense on its own, and together they add up to the wall of panels that scared me off in the first place.

The trade-off: every feature I add makes my studio a little more like the thing I couldn't learn. Keeping it simple is a decision I have to keep making, not one I made once.

Let a budget shape the architecture

I wanted it to run on my flat-rate Claude Max plan, not a pay-per-token API key. It turned out a subscription login only works through Claude Code itself. So instead of fighting that, I designed around it: the studio launches Claude Code behind the scenes and hands it the studio's tools, while the chat stays inside the app.

The trade-off: a slightly unusual setup. In return, every video costs me nothing extra beyond the plan I already pay for.

Remove a limit that protected nothing

Early on, the agent had a cap on how many steps it could take per request, and long builds kept dying halfway. I asked myself why the cap existed at all. On a flat-rate plan it saved no money, and it threw away finished work. So I removed it by default and added a Stop button instead.

The trade-off: a runaway turn is possible, so I'm the one who stops it. In return, the agent finishes what it starts.

One agent per video, not a team per video

My first plan was a lead agent splitting one video into sections for several workers. Then I looked at what I actually make: short videos. Splitting a 30-second clip adds seams, a shared camera and messy undo for almost no speed gain. Running several agents on separate videos gives the same output with none of that.

The trade-off: a single long video doesn't get faster. In return, the system stays simple.

Do the math before buying the feature

I wanted music on every video. My music plan gives about 33 minutes of generation a month, and six agents making two videos a day would need around 180. Fresh music per video was never going to work. So I generated a library of 16 tracks once, all built with the same structure, and each video borrows the stretch that ends on the track's own outro. Then I cut scope again: only YouTube gets library music, because on Instagram and TikTok I add trending sounds myself for reach.

The trade-off: less variety than fresh tracks. In return, music fits the budget and actually helps where it's used.

Keep the risky parts human

The agents can research, build and render on their own, but they can't post. Auto-posting would mean developer apps and review processes on every platform, and one bad autonomous post is a brand risk I don't want. Every fact also needs a source, and agents only use material we own.

The trade-off: I'm still in the loop for every post. That's the point.

Read the license before the README

I turned down a strong video-tracking model because its license didn't allow commercial use, and used a different approach instead. I also replaced several libraries with small in-house code to keep dependencies down, and noted one license (in beat detection) that would need a second look if I ever sell the studio. Licenses are product decisions, not paperwork.

What surprised me (and what broke)

The agents all had the same idea. All six picked "by the numbers" niches, and three of them landed on money and prices. I'd only told partners on the same platform to stay different from each other. Lesson: if you want variety, it has to be a rule for everyone.
The toolbox shapes the content. The chart tools were the strongest, so the agents made chart videos. What you build into the tools quietly decides what gets made.
I'm the bottleneck now. All 25 videos are sitting in my review queue. The work moved from making to judging, which is exactly what I wanted, and also a little humbling.
The agents are decent coworkers. One left a note saying the critic seemed to mishandle multi-line text. It was right, and I fixed it.

Real use also broke things that tests never did. When my laptop went to sleep, the agent's background browser died silently and previews froze. And some of the generated music ended with up to three seconds of silence, which would have ended videos on dead air. Both are fixed now.

The part I can't automate

I've started posting these videos on YouTube, Instagram and TikTok. Part of it is a real experiment: will anyone be able to tell that AI did most of the work? Part of it is just wanting the channels to grow. I don't have numbers worth sharing yet, so check back in a few months.

But I've already learned the uncomfortable lesson. Building the studio was the easy part. Getting people to actually see what it makes is the hard part, and I'm not sure anyone ever fully solves it. Coding can be delegated. Distribution, so far, can't.

What I learned

I built this in about two weeks with Claude Code writing most of the code. My job was the design, the rules, the trade-off calls and actually using the thing until it broke. Some of the best features came from frustration, like autosave, which showed up right after I asked, "so the previous work never got saved?"

The biggest surprise was After Effects itself. I started out thinking it was just overcomplicated. After watching my own tool grow one sensible feature at a time, I get it now. Every panel that intimidated me was probably someone's very reasonable product decision.

The biggest lesson: AI didn't remove the work, it moved it. I still don't know After Effects, and now I don't need to. My time goes into deciding what's worth making.

Honestly, that's the part I wanted to be doing all along.

← All posts