Edit Video with GPT-6 Astra · The /animate Skill in 3 Steps

Quick Summary

• Jay E of RoboNuggets demonstrates his /animate skill: he gives GPT-6 Astra raw recordings of eight scripts and gets eight vertical edits with cuts, captions, motion graphics and visual assets. This demonstrates his configured workflow, not guaranteed results for every recording.
• The difference between a generic skill and a refined one lies in the persistent instructions: visual identity, layout choices, readability on mobile, animation types and examples of reusable elements.
• The video method has three steps: getting to know the tools, reviewing the first outputs with specific feedback and asking the agent to update the skill itself. Then the cycle is repeated for each project.
• In Jay's implementation, AssemblyAI interprets speech, HyperFrames composes and renders video from web elements, and GPT-6 Astra coordinates the process. Whisper and other platforms are mentioned alternatives, not skill requirements.
• The PDF provided is called Build Your Own /generate Skill and is nine pages long. It teaches you how to forward image and video requests to APIs, save references and outputs, estimate costs and create logs. It is not the guide/animate referred to in this video.
• PDF rules are useful as a complementary layer: /generate can produce assets for a library that /animate uses in assembly. The article clearly distinguishes the two functions and includes original prompts for testing the cycle.

The result: eight Reels from one recording

At the beginning of the video, Jay presents a long recording in which he reads eight scripts. Ask GPT-6 Astra, in the ChatGPT application, to use its /animate skill to produce eight vertical videos. The sample shows cuts, text graphics, browser mockups, captures, and 3D objects. Credit for recording, workflow and demonstration: Jay E and RoboNuggets.

The phrase a single prompt describes the user interface, not the amount of work behind it. The skill already contains repeatable decisions, previously designed elements and rules that Jay added. A first request without these references produced, according to him, an acceptable but generic video.

The practical conclusion is to make editorial decisions reusable. The agent may perform time-consuming operations, but the person continues to define what deserves emphasis, what the brand should look like, and what result is good enough to publish.

Generic skill versus refined skill

Both receive raw video and return edited video. The difference is in the quality of the instruction system. The generic version can always choose the same type of frame, fill the frame with secondary text or apply a template aesthetic. The refined version incorporates corrections taken from real examples.

Jay suggests starting with a functional skill, generating one or two pieces and evaluating the result as if you were giving notes to an editor. Feedback must be observable: when a number appears, use a large leading number; shorten URLs in mockups; alternate compositions; capture interfaces with proportion and typography suitable for the vertical format.

This process does not automatically create quality. Makes criteria explicit and verifiable. Even after the skill improves, each video needs review of cuts, synchronization, subtitles, media rights, factual coherence and export.

Step 1: Know which tools are included in the pipeline

For speech, Jay uses AssemblyAI as a transcription service. It also mentions Whisper as a local option and ElevenLabs Scribe as an alternative. Speed ​​and cost comparisons are numbers reported in the video, sensitive to model, hardware and price; consult each supplier before using them in a quote.

To compose the video, it uses HyperFrames. The official project describes an open source engine that transforms HTML, CSS, media and controllable animations into MP4. This allows the agent to build overlays, graphics, transitions and browser layouts without relying on a manual timeline for each element.

GPT-6 Astra coordinates the tasks in the example. The video also mentions Claude Code as an alternative agent. The choice of template does not replace the two operational capabilities: interpreting the source material and producing a video composition that can be previewed and rendered.

For a first implementation, it separates four states: preserved raw file, revised transcription, editable composition project and final export. This separation helps correct a legend or graphic without destroying the source.

Step 2: Refine with visual direction and structured feedback

The first layer of refinement is the visual direction. Jay shows the Refero Styles catalog and its Markdown style files as a starting point. Then adapt colors, typography and elements to RoboNuggets' visual system. A design system from another company may inspire hierarchy or composition, but copying the identity entirely creates problems of differentiation and rights.

The second layer is concrete rules. Instead of asking to make it better, point out the timecode, the fault and the desired fix. Example: at 00:12, the URL is cut off; use short hold and a vertical catch. At 00:19, the number is important; replace the small text with a main counter. At 00:27, the same mockup is repeated; alternate with a diagram.

The third layer is an element library. In the video, Jay brings together 3D graphics, browser layouts, animated entries, infographics, images, icons, logos, backgrounds, effects and music. The agent receives the paths of approved assets to reuse them. A library with defined versions, licenses and proportions produces more consistency than generating everything again for each Reel.

Step 3: update the skill, then test again

After commenting on the first versions in the same session, Jay asks Astra to calibrate the /animate skill with this feedback. The instruction transforms specific notes into rules for the following process. The stage is only complete when a new clip shows that the rules work, without ruining what was already good.

A useful review keeps three things: the original example, the requested change, and the version of the skill that incorporated it. If a rule only makes sense for one client or format, it marks its scope. Otherwise, a preference from one project could contaminate all future videos.

The recommended cycle is short: render a sample, review on desktop and mobile, give accurate feedback, update the skill and test another recording. Only then is it worth launching eight batch exports.

The attached PDF is about /generate, not /animate

RoboNuggets' Build Your Own /generate Skill document is nine pages long and has already been analyzed in another JQ VISIONS article about Jay's alternative to Higgsfield. This is the file provided by the user here. The current video mentions a separate PDF for /animate and element prompts, but the document you received does not contain such a guide. It would not be correct to present your texts as the exact editing recipe shown in this tutorial.

Pages 1 and 2: /generate is a skill for requesting images and videos from external models. The proposed flow has four operations: choosing model and supplier, preparing real references, calling the API and recording the result. The suggested structure includes SKILL.md, one recipe per model, the generations folder, and a refs subfolder. This is asset generation, not cutting and assembling a recording.

Pages 3 to 6: The guide shows how to save Kie AI, fal.ai and WaveSpeed AI keys in an .env file, define routing rules in SKILL.md and document endpoint, authentication, cost, request and response format in each recipe. PDF templates, IDs, and currency values ​​reflect the date of the document and require current verification before a paid call.

Pages 7 and 8: before paid video, the agent must declare expected cost and wait for explicit authorization; image drafts should start with an economical option; logos and faces must be included as reference files. The document also recommends a JSON next to each output, with prompt, model, parameters and date, and describes a local gallery to consult all resources.

The connection between the two systems is simple: /generate can create an approved illustration, background or object, save the provenance and deliver it to the element library; /animate chooses where this asset goes, composes the video and renders the final version. This integration is an editorial proposal, not a function demonstrated between the two skills in the video.

Practical blueprint for your first version

1. Choose a short recording and format, for example 9:16 with subtitles. 2. Defines the transcription rule and mandatory review of names, numbers and brand terms. 3. Gives the agent its own design system with fonts, colors, safe margins and text limits per frame.

4. Asks the initial skill to plan beats, extract clips, choose only the necessary elements and assemble an editable composition. 5. Confirm rights to music, images, logos, screenshots and voices. 6. Render a short sample and note three to five specific corrections.

7. Update the skill only with rules that must survive the project. 8. Test a second recording. 9. Store assets and logs so that a review can reproduce decisions. 10. Only then automate batches. Gain should be measured by time to approved video, not time to first MP4.

Verdict: Automate execution, preserve direction

Jay's demonstration makes much faster editing of educational videos with recurring patterns plausible. The most interesting contribution is not the model name, but the working memory: what an editor would do, what the brand allows, and how to check the result.

PDF /generate reinforces this idea at another point in the pipeline. Cost rules, real references, recipes per model and logs prevent asset creation from being a black box. For those who work with video every week, the two skills can complement each other, as long as generation and editing are not confused.

If you're looking for the specific /animate guide mentioned by Jay, check the current video description or the RoboNuggets community. The file provided for this article is the /generate tab, and has been treated as such.