This Just Killed AI Video Editing With GPT 6 Astra

AAI LABS
Computing/SoftwareAdvertising/MarketingInternet Technology

Transcript

00:00:00a lot of people have been letting models make their videos from start to finish from planning
00:00:04the shots to editing them together and adding the music and knowing how to build these videos
00:00:09matters even more if you're building a product because these videos are the quickest way to
00:00:13draw people's attention to what you've built without having them actually use it but if you
00:00:18dive straight into creating the video you won't get the quality you want and you'll also waste a
00:00:22lot of time and tokens to avoid that you need to follow a whole workflow that makes sure that the
00:00:27video turns out just how you want it for our workflow we use gpt6 astra inside cloud code
00:00:33since astra is one of the best models at video editing and this setup lets us use the strengths
00:00:37of both so if this is your first time here we're a software company and this is ai labs and in this
00:00:42video we're going to go through that workflow step by step including what problems you might run into
00:00:47along the way and how to fix them now you might be wondering why we're running astra inside cloud code
00:00:52when astra already comes with its own tools in codex the reason is that we prefer the tools cloud code
00:00:58gives us and that's why we use it for pretty much everything we build and the reason we're not using
00:01:02cloud's own models inside it is that astra generally has a better design direction and it also came out
00:01:08ahead on design when we compared it with fable in our previous video so with this setup astra provides
00:01:14the creative direction for the video and cloud code runs the tools to create it but if you don't want to
00:01:18use cloud code you can also use codex directly or any other coding agent you like if you're using
00:01:24codex you can skip the part where we set up astra in cloud code and go straight to installing the video
00:01:29tools so if you want to use estra in cloud code like us you have to take some steps first you need to
00:01:35set estra up so that cloud code can recognize it as a model one way of doing that is paying for openai's
00:01:42api and connecting that to cloud code but we wouldn't recommend that because the model is expensive and
00:01:47creating videos uses a lot of tokens that will get really costly for you the other way is using astra
00:01:53through the codex subscription you already have but to make that available in cloud code you need to
00:01:58use a tool called cli proxy api it runs a small server on your computer that connects cloud code to
00:02:04the coding agents you've already installed that server lets you use your existing subscription without
00:02:09paying separately for the api it works with a lot of agents and we use this tool in our kimmy design
00:02:14video as well to use it you first need to install the tool by running the brew install command in your
00:02:20terminal once it's installed there's some setup to do which includes changing a few of its settings
00:02:25and making an api key for it but instead of doing all of that yourself it's way easier to just ask
00:02:30whichever agent you're using to set it up for you and mention in your prompt which model you want to
00:02:35use through the tool once the agent has finished those settings it gives you the api key that cloud code
00:02:40will use to connect to this tool you need to save that key for the connection step once this tool is
00:02:45installed it will be started automatically you don't need to run any command to start it now the
00:02:50key the agent gave you isn't the normal kind of api key that charges you for every token this key
00:02:55just lets you use the coding agents you've installed through a server running on your computer it only
00:03:00works on your system and not on anyone else's next you need to connect cli proxy api to your codex
00:03:06account using its login command when you run it in your terminal it opens the browser where you sign
00:03:11in with your account once you log in the tool saves your codex login so that it can use your plan
00:03:16so once you're logged in you need to launch claude code with astra now normally claude code sends
00:03:21everything you ask it to anthropic which is how you get access to claude's models but you don't want
00:03:26claude's models right now so you need to change a setting called the base url which makes claude code send
00:03:32everything to the server running on your computer instead then you add the api key that you copied
00:03:36from earlier and type in gpt6 extra as the model and once you launch claude code gpt6 extra will be
00:03:43available in it and you can start using it and once that's done there's one more step for the setup which
00:03:48is installing the tools that help with the video creation process you can pause here take a screenshot
00:03:53of this list and give it to your agent and it will install all of them for you those tools come with
00:03:58skills that tell the agent how to use them when making the video and once they're installed you're
00:04:03ready to start creating the videos but before we move on with the workflow it would be great if you
00:04:08subscribe to the channel and hit the hype button this small gesture of support goes a long way for us
00:04:13for our product video we used an artistic clothing store where you can put together an outfit and buy
00:04:18it we built the store using the app template we have already shared in our community which gives you a
00:04:23starting point for building your own app before creating the video we needed to plan which parts
00:04:28of the store to show and how to show them planning is really important since creating the video can
00:04:33take hours if you don't check the plan first you could spend hours creating a video that isn't what
00:04:38you wanted for planning we gave claude code a prompt with the address where our store was running so it
00:04:43could open the store and plan the video around it we also included the length and layout we wanted we
00:04:48wanted a 30 second video in a 16 to 9 layout which is the horizontal layout you see in product demo videos
00:04:55you can choose the 9 to 16 if you want the vertical video for short form content we wanted music in
00:05:00our video without a voiceover so we included that in the prompt too now if the plan only describes the
00:05:05shots in text it's hard to picture what the video will look like so we asked it to show the plan on
00:05:10our browser so that we could see what the video would look like and ask for changes we also asked for a
00:05:15place to leave comments on the plan so that we could tell the agent which parts we wanted changed we
00:05:20also asked it to use anthropics and linears product launch videos as references for how our video should
00:05:26look and move it's important to give references because that way the agent knows what idea you
00:05:31have in mind the references also help it plan how the view moves across the page and zooms in on a
00:05:36button once we gave it the prompt it loaded the skills that got installed with hyperframes hyperframes is
00:05:41one of the tools we installed to help the agent create the video the agent used its video skills
00:05:46to plan the shots then use the other skills to plan out the whole video and once it was done it opened
00:05:52the whole plan on the browser the plan showed what would appear at each second of the video how the
00:05:56camera would move and what would happen on the site and just like we asked there's a comment section
00:06:01where you can tell the agent what you want changed you can add many comments as you want to ensure
00:06:06that the plan is to your liking once your review is done you can ask it to check the comments update the
00:06:11plan and make the changes once you're satisfied with the plan you ask the agent to lock the plan
00:06:16so that the agent knows that this direction is the final one after that you can move to the next
00:06:21step but first let's have a word by our sponsor luma ai every ai video tool makes you pick two out of
00:06:27three fast cheap good ray 3.2 doesn't follow it we ran cinematic shots and anime through it and everything
00:06:34came back native 1080p full hdr four times faster and a third the cost of ray 3. the speed side by
00:06:41side was almost annoying we'd hit generate expecting to wait and it was done what surprised us more was
00:06:46accuracy prompts landed closer to what we asked for glitches dropped off and it held a consistent style
00:06:52frame to frame if you've ever rendered something and thought close enough i'll patch it in post you
00:06:57don't need to anymore ray 3.2 supports keyframes and a modify tool so you can go back into a finished clip
00:07:04and fix it luma now also hosts c dance 2.5 a big step up from c dance 2.0 plus gen 4.5 and cling 3.0
00:07:13so you can pick the best model try ray 3.2 links in the pinned comment or scan the qr code on screen
00:07:20the next step is basically creating the shots separately at this stage we're creating a basic
00:07:25version of each shot so we can check the camera movements before spending time polishing it the
00:07:30reason we create each shot separately is that if one shot isn't to your liking you only need to make
00:07:35changes in that one shot rather than having to redo the whole video again this is also the point where
00:07:40you check the zooms and the other camera directions to see if they're the way you want for that we gave it
00:07:45a prompt telling it to open the browser record every shot at full screen resolution and turn each one into a
00:07:51video so that we could check whether the camera movements were done correctly after our prompt it started
00:07:56working on the shots it created the first one and asked us to review it before moving on so that if
00:08:01something was off we'd catch it early instead of finding the same problem repeated throughout the video
00:08:06it then showed us the images and recorded clip in our browser if you want changes to the first one you can ask for them
00:08:12and once you're happy with it you can tell it to generate the rest when it's finished you can review
00:08:17everything together in your browser in our case a few clips were really shaky for some reason and
00:08:22some of the zooms weren't the way we had described in the plan so we asked it to fix those and you can
00:08:26keep requesting changes to any part you don't like until you're happy with the final result the reason
00:08:31we check these basic shots first is that if a shot needs to be redone we haven't already spent time
00:08:36polishing it once we've approved what each shot shows we can improve the movement and combine the shots
00:08:41into one video at this point we had separate clips with basic movement and zooms we needed them to
00:08:47be combined into one video that flowed smoothly to polish them we asked claude code to improve those
00:08:53movements and combine the clips so that each shot moves smoothly into the next in that prompt we
00:08:58explained how the camera should move where it should zoom in and when the typing should happen creating
00:09:02the combined video took a lot of time and once it finished we had the shots combined into one video
00:09:07without music the agent saves the video in the folder it's working in so you can open it from there and
00:09:13check it if something doesn't look right you need to tell the agent where it happens in the video
00:09:17in our case it placed one of the shots too far to the right so we asked it to move that shot back
00:09:22into place then once all the changes have been made you need to ask it to review the video again and
00:09:27check for any issues that are still there the tools we installed already come with checks that look for
00:09:32small issues in the video so when those run the end video does not contain any of those issues the
00:09:38final step for the video is to add in music it's a really important step because the music is what
00:09:43keeps the viewer hooked through every shot and it makes the video a lot more engaging to watch if you
00:09:48just ask the agent to add music it will pick a track by itself but that track might not fit the video
00:09:54you need to give it some guidance on the kind of music you want so to guide that we created a skill
00:09:59that basically tells the agent exactly how to pick the music the skill reads the product style from the
00:10:04plan and checks how often the video changes from one shot to the next it uses that timing to look for
00:10:10music with a beat that fits the video it searches a site called mix kit for free music in categories like
00:10:16calm and piano to create the skill we gave claude code a prompt telling it the exact workflow it needed
00:10:21to follow and it created the skill with everything it needed to find the music and add it to the video
00:10:26if you want this skill too you can copy the prompt and give it to your agent so that it follows the
00:10:31same workflow and creates the skill for you as well to use it you just need to type slash and the skills
00:10:36name and it finds the music and adds it to your video it then creates a new version of the video with
00:10:42the music added now we have refined the whole workflow you just saw in the form of the skill that you can
00:10:52use directly in your workflow that skill has been made with multiple rounds of testing and refining and
00:10:57it is available in ai labs pro which is our community so if you found value in what we do and want to
00:11:03support the channel this is the best way to do it the links in the description that brings us to the end of
00:11:08this video if you'd like to support the channel and help us keep making videos like this you can do so
00:11:13by using the super thanks button below as always thank you for watching and i'll see you in the next one

Key Takeaway

By routing GPT-6 Astra through Claude Code via CLI Proxy API and using Hyperframes skills, developers can build fully automated, interactive 30-second product demo videos without paying per-token API charges.

Highlights

  • Running GPT-6 Astra inside Claude Code using CLI Proxy API avoids direct OpenAI API token costs by leveraging an existing Codex subscription.

  • A custom script or proxy server connects Claude Code to third-party models by overriding the local base URL setting.

  • Drafting an interactive video plan on a local web server with comment capability prevents wasted render time and unnecessary API token usage.

  • Installing Hyperframes skills allows AI coding agents to control browser camera movements, zooms, and screen captures frame-by-frame.

  • Automating background music selection involves searching Mixkit for free audio categories aligned with product visual transition timing.

Timeline

Configuring GPT-6 Astra inside Claude Code using CLI Proxy API

  • Direct OpenAI API usage for AI video generation incurs high token costs.
  • CLI Proxy API creates a local server on the host machine to route Claude Code requests through an existing Codex plan.
  • Overriding the base URL in Claude Code directs traffic away from Anthropic servers to local models.

Creating video content requires substantial token consumption. Utilizing CLI Proxy API allows developers to bypass standard API token charges by leveraging existing subscription tiers. Setting the base URL in Claude Code points internal requests to the local proxy server, making GPT-6 Astra available directly inside the editor environment.

Interactive shot planning and web-based feedback

  • Unplanned video generation leads to hours of wasted rendering time.
  • Hyperframes skills enable AI agents to plan camera paths and display interactive blueprints directly in the browser.
  • Locking the final storyboard prevents unwanted modifications during subsequent shot production.

Before generating raw video frames, the agent maps out shot timing, aspect ratios, and camera focal points based on direct product URLs and reference videos like Linear or Anthropic launches. Displaying this plan on a local web page allows creators to leave exact feedback and request structural updates before locking the storyboard.

Luma AI Ray 3.2 overview

  • Luma AI Ray 3.2 renders native 1080p HDR video at four times the speed and one-third the cost of Ray 3.0.
  • Keyframe controls and modify tools allow direct fixes to rendered clips without starting over.

Generative video production relies heavily on rendering efficiency and consistency. Ray 3.2 improves prompt accuracy while maintaining frame-to-frame stability across cinematic and styled renders. Built-in editing tools let users modify post-render frames directly.

Generating and refining individual video shots

  • Rendering shots separately isolates errors and speeds up revisions.
  • Reviewing the initial shot catches movement and zoom issues early in the pipeline.

Generating each scene as an independent clip prevents cascading rendering errors across the sequence. Reviewing early shots identifies camera shake, incorrect zoom levels, or layout bugs so adjustments can be made before full composition.

Combining scenes and adding synchronized audio

  • Custom skills analyze visual transition timing to select matching audio from Mixkit.
  • Automated video checks eliminate small visual artifacts before final export.

After scene capture, Claude Code stitches the individual video segments into a single cohesive video file. A dedicated audio selection skill evaluates transition pace and product style, automatically querying royalty-free music libraries like Mixkit for fitting backing tracks.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video