How I make content with Astra

I've been using Astra in Codex to edit videos we've shipped on the official ChatGPT Instagram. I've also been using it to pull clips from longer videos and create product animations.

And what's really impressed me is is the editing taste that Astra has without my explicit direction. I'll give it the assets, tell it the story I want, and let it work through things like take selection, pacing and captions. I want to show how I've been doing this, because a lot of it is very possible right now and makes creating things a lot easier.

UGC videos

For UGC-style, vertical, face-on content, I like to go to a place outdoors with nice lighting. And preferably some greenery. Then I'll film clips in succession for what I imagine the video will be. This could be me talking to the front-facing camera on my iPhone, or taking a video of another phone to show what's happening.

When I'm recording, I'll usually do enough takes to get a good one. I'll say “cut” between takes, and “that's a good one” after a take I like. Sometimes I'll say what the beat is at the top of the recording.

I say these things because I treat Astra like an editor. When it's processing the clips, it's reading the transcripts and looking at the frames, so it can “see” and “watch” the footage. Those hints guide it to the takes I want.

Once I have the takes, here’s what I do:

From there, Codex gets transcripts through OpenAI's transcription API and breaks the clips into frames Astra can inspect. Astra uses that information, along with my direction, to figure out which takes to use and how to piece the video together. The word timings also help it put the captions in the right place.

Codex uses FFmpeg to assemble the edit. And Astra has a great sense for pacing. The Odyssey video is a good example of this. I gave it the story I wanted, and it figured out how to make the cuts feel crisp, make the captions look good, and emphasize the right moments.

Astra also spends time reviewing its work, going back through frames and audio. I'll often have it use subagents for this. I use a gauntlet loop: have a fresh reviewer look at the output, make the changes, and review it again. This can take a while, especially for things like blurs that need serious attention to detail.

One of my favorite examples is the catch-up-on-work video. Around 30 seconds, the blurs follow the text as I scroll. You can still see what the product is doing, while the sensitive information is obscured.

When Astra is happy with its work, it gives me the video to review. There are usually some iterations. I might want a different title, a tweak to the UI, or a little more time on a particular shot. I'll point to the timestamp, explain what I want changed, and tell it what to leave alone.

Something I've been doing is asking Codex to spin up a Frame.io-style review interface in the in-app browser. That makes it easy to leave comments on different parts of the video, which is a familiar way for me to give feedback on an edit.

And I would give you a framework for the Frame.io-style editor, but I don't use one. I just ask Codex to spin one up, and it's nailed it quickly each time I've done this. Create a skill for this if you feel compelled.

I'll upload the clips, give my direction and then let it cook for a while. Because it does take a while. The time I'm putting into reviewing is more like three minutes here, three minutes there, checking what it's done and leaving a note.

So I can spend about 30 minutes getting clips for a video, then around ten minutes reviewing the edits across those short sessions, and get it edited that same day. Astra is working for longer than those ten minutes in the background.

Something I've liked doing is running these edits on dev boxes. The iterations can take some time, and I don't like relying on my laptop staying on for progress.

Clips from longer videos

I also use Astra to pull clips from longer videos. It works through the transcript and looks at the frames, finds good moments, and cuts them so they make sense on their own. It produced five clips from a roughly 44-minute podcast this way.

And I can ask for the sizes and formats I need for different platforms, with captions and framing adjusted for each.

Motion design

When I'm creating motion design videos, I'm trying to show a process more simply. And so there's some product and viewer intuition you need. You have to figure out what somebody actually needs to see to understand what's happening, and how to show that without making them follow a whole screen recording.

1. Give it the product context. I'll give Astra whatever context I have for this. It could be a recording of me doing the thing, or screenshots, or another Codex thread where I've already done it. For the Spotify video, I had a thread where Codex was acting as a DJ for me. I gave it that thread and asked it to figure out how to explain what was happening.

Computer use is useful here because it can go look at the product. Astra is also great at recreating things from screenshots. But the most useful context, when I can give it, is the actual codebase. It can read how the interface is put together and use the real components and icons. There's nothing quite like it having that to work from.

2. Get the still right. What I like to do is get a still that looks good first. I'll work with it on the layout, what we're showing, how much text there is. And I'm pretty particular about the product being accurate. If we're showing an interface, I want it to look like the actual interface.

3. Bring it to life. Then, once I like the still, I'll ask it to bring it to life. Have the prompt type out, have the cursor do the work, bring the different windows into view when they're needed. I give it an idea of what I want to happen, but I let it work out a lot of the timing.

And what's been interesting is how it chooses to build these. For the Spotify one, it used ImageGen for the background and React, HTML and JavaScript for the interfaces and animation. We used to use Remotion for some of this. This one didn't need it. I'm more interested in whether it looks good and communicates the process than which framework it ends up using.

An animation of the workflow, with compressed timing, rather than a continuous recording of the task running.

4. Review the details. I'll have it use the same kind of review process here. There are a lot of little things to check: whether you have enough time to read something, whether the cursor is going to the right place, whether the thing we're showing is actually what the product does. In the Spotify video, a review caught that the queue action was too small to read, so it added a zoom and held there long enough for you to see it. That's the kind of detail I want it working through before it comes back to me.

I also use Astra for the thumbnails across all of these videos.

If you want to try this, start with some footage you already have. Give Codex the clips, talk through the story you want to tell, and let it cook.