How to Edit Videos With GPT-6 Astra and Codex: A Practical Guide

For creators, YouTubers, educators, and small production teams. Documentation checked September 4, 2026.

Astra can help you turn a folder of footage into an edited video. You describe the result, give it access to the source files, and let Codex use software to inspect, cut, arrange, caption, and export the material.

The result depends on the tools available in your session. Astra supplies the reasoning and instructions; Codex provides the working environment; an editing tool or rendering framework produces the video file.

That distinction matters. OpenAI lists GPT-6 Astra as accepting text and images and producing text. Its supported tools include shell execution, computer use, and integrations. There is no native MP4 output modality listed for the model itself. The workflow in this guide applies those capabilities to editing through tools. Selecting Astra alone does not configure a video studio. OpenAI’s Astra model specification.

For a first project, start with something specific: shorten a talking-head recording, turn a screen demonstration into a vertical clip, or add branded titles to existing footage. You can then build a repeatable workflow around the parts that work.

Evidence note: This is a documentation-based how-to guide. During preparation, the two FFmpeg editing recipes below were run on synthetic footage using FFmpeg 8.1.1 on macOS. Their output dimensions, duration, codecs, and full-file decoding were checked. The longer creative workflows and prompts are examples to adapt, not a completed Astra editing benchmark or a claim that every editing application was tested.

1. What Astra and Codex each do

Think about the work in four parts:

A good request connects all four. “Edit this video” leaves almost every decision open. “Make a 60-second explanation from these two files, preserve the caveat, add captions, and export a vertical MP4” gives the agent an outcome it can work toward.

OpenAI describes Astra as a model for complex work across code, apps, and research. Those general capabilities explain why it can participate in an editing workflow; they do not establish a measured improvement in editing speed or quality over another model. OpenAI’s model guide.

What is reasonable to ask it to do?

With the appropriate tools, you can ask Codex to:

• Inspect files and report their duration, dimensions, frame rate, and audio tracks.
• Assemble clips from specified source timestamps.
• Draft a cut list from a timestamped transcript.
• Resize, crop, pad, or reframe footage.
• Prepare captions and place them on a timeline.
• Add titles, logos, screenshots, or animated graphics.
• Make several versions from an approved project.
• Render output files and check their technical properties.

Treat footage selection, automatic reframing, and caption accuracy as reviewable decisions. A centered crop can remove the product being demonstrated. A transcript can mishear a model name. A tighter edit can accidentally remove the sentence that qualifies a claim.

For detailed color matching, complex sound repair, or a long documentary with hundreds of subjective decisions, a conventional editor may still be your most efficient finishing environment. That is an editorial recommendation about workflow fit, not a tested limit on Astra.

2. Choose an editing route

You do not need every tool in this table.

FFmpeg documents conversion and filtering through its command-line tool. Remotion describes video as React components rendered over time. HyperFrames documents importing, placing, trimming, and reviewing existing media. These are different ways of turning the agent’s decisions into actual frames. FFmpeg overview, Remotion fundamentals, HyperFrames media guide.

My suggested starting point for a creator is one short local edit. Use FFmpeg for a straightforward cut. Choose Remotion or HyperFrames when a reusable visual layout is central to the result.

3. Set up Codex and select Astra

OpenAI’s current documentation places Codex inside the ChatGPT desktop app. If an older tutorial tells you to open the separate Codex app, the visible branding and navigation may differ from what you see today. The current setup is to install the desktop app, sign in, choose a working location, and select Codex. Desktop app setup.

Use a local folder for the workflows in this guide. A cloud task should not be assumed to have access to footage or installed apps on your computer.

1. Open the desktop app and select Codex.
2. Open a folder for the video project.
3. Use the model control beneath the composer to select GPT-6 Astra if it is available.
4. Begin with the default reasoning effort. Raise it only when the planning or debugging warrants it.
5. Give the task access to the files and tools required for this edit.

Model access depends on rollout, account, sign-in method, and client. Some pickers expose a Power control and an Advanced menu. If Astra is missing, check the available models and workspace access instead of assuming a configuration change can grant access. Enterprise workspaces may also require an administrator to enable it. Model selection and availability, Enterprise model access.

These workflows are also useful with another capable Codex model. Astra is the focus here, but the renderer and source files remain essential whichever model you use.

Start with a capability check

Paste this before requesting a large edit:

I want to edit local video files in this project using Astra and Codex.

Inspect this task's available tools and the files in source/.
Report:
- Which source files you can access.
- Which local editing and rendering tools are installed.
- Whether transcription is available and whether it runs locally or
  sends audio to an external service.
- Whether you can inspect extracted frames and preview the output.
- Any missing dependency that blocks a first short edit.

Recommend one editing route for these files.
Do not install software or start a paid service during this inspection.

The useful answer names real tools and real files. A generic assurance that “I can edit videos” is insufficient.

If FFmpeg is missing, start with its official download page, which links to packages and compiled builds. Ask Codex to identify a suitable build for your operating system and the features needed for your edit, then review the installation it proposes. Remotion and HyperFrames setup are covered below.

For a command-based setup, Codex can inspect ffmpeg -version, ffprobe -version, and the installed encoder and filter lists. For a framework, it should check that framework’s dependencies and installed version.

A successful installation check does not prove that every feature exists. The FFmpeg installation used while preparing this guide could trim and resize video, but ffmpeg -h filter=subtitles reported that the subtitle filter was unavailable. FFmpeg’s documentation specifies a libass-enabled build for that filter. FFmpeg subtitle requirements.

4. Give the project useful source material

A simple folder arrangement makes the work easier to follow:

kingy-video/
  brief.md
  source/
    presenter.mp4
    screen-demo.mp4
  assets/
    logo.png
    approved-music.wav
  transcripts/
  project/
  work/
  outputs/

This is a suggested organization, not a required Codex format. Use work/ for temporary files, project/ for editable source, and outputs/ for deliverables.

Put the footage inside the folder before starting, or explicitly identify another accessible location. If files sit on an external drive, say which drive and verify that the task can read it.

The brief should answer six questions:

• Who will watch?
• What should they understand?
• Which footage should be used?
• How long should the result be?
• Where will it appear?
• What must remain accurate or intact?

For a Kingy.ai product explanation, the required elements might be a visible demonstration, the relevant limitation, and a short closing direction. Include correct spellings for product names and model names. Supply exact wording for any claim that must appear on screen.

Provide music, logos, fonts, and footage you are entitled to use. If the recording contains unreleased material or private information, decide which services may receive it before choosing a transcription or generation route.

Local rendering does not mean the entire AI workflow is offline. OpenAI distinguishes the command sandbox’s network controls from model requests, connectors, browser activity, and other service connections. Codex network and permission boundaries.

5. Make your first complete edit

A good first assignment is a short explanation assembled from a presenter recording and a screen demonstration.

The following example asks for a 60–90-second landscape video. Change the filenames, duration, and destination to match your project.

Prompt: create a reviewable rough cut

Create a 60–90-second rough cut from:
- source/presenter.mp4
- source/screen-demo.mp4

Audience: creators deciding whether the demonstrated tool is useful.
Message: explain the workflow, show the result, and preserve the main caveat.
Format: 1920x1080 landscape. Inspect the sources and choose an appropriate
frame rate; explain any frame-rate conversion.

First inspect file metadata. Use an available authorized transcription
method for dialogue, and inspect timestamped frames where visuals matter.

Build the edit around:
1. A clear opening statement.
2. The most useful demonstration.
3. The result and its limitation.
4. A short closing direction.

Keep original speech and natural pauses where they help comprehension.
Do not invent quotes, results, or product capabilities.
Use only the supplied assets.

Save an editable project or reproducible script, a cut list with source
timestamps, and outputs/rough-cut-v1.mp4.
Render a short representative test before the complete rough cut.
Preserve the source files. Report anything you could not verify.

This gives Codex permission to produce the local edit and defines what “done” means. If it needs an unavailable tool, it should identify that dependency clearly.

The rough cut should resolve structure before detailed decoration. Watch the opening, the transition into the demonstration, and the final caveat. If the argument is wrong, animated captions will not fix it.

Transcript timing needs a clear reference

A timestamp such as 00:42 can mean 42 seconds into the original recording or 42 seconds into the edited output. Those are different locations once material has been removed.

Ask for source timecodes in the cut list and output timecodes in feedback on a render. A useful cut list records:

The agent must fill those cells from the actual material. Do not accept invented precision.

A plain transcript is useful for understanding the argument. Word or segment timing is needed for precise editing and captions. OpenAI’s transcription documentation identifies whisper-1 as supporting its timestamp-granularity parameter; the transcription tool is separate from Astra itself. A different installed transcription tool can be used if it provides the timing you need. Speech-to-text timestamps.

Prompt: tighten a rough cut

Revise outputs/rough-cut-v1.mp4 using the saved project.

All timestamps below refer to the output video:
- 00:00–00:07: reach the main point sooner by trimming the lead-in.
- 00:18–00:25: keep the demonstrated setting visible while it is explained.
- Near the ending: preserve the full caveat and let the last sentence finish.

Improve these sections and save rough-cut-v2.mp4.
Keep the approved structure elsewhere.
Update the cut list and identify the source changes you made.

Specific feedback reduces unnecessary reworking. “Make it more professional” gives the model little guidance about whether the problem is pacing, framing, typography, or sound.

Make dialogue easy to hear

Audio deserves its own decision. A quieter music bed, a repaired level change, and a clean ending may do more for an explainer than another visual effect. Keep an unprocessed copy so you can compare any cleanup.

Improve the audio in the current rough cut using available local tools.

Prioritize clear dialogue. Inspect the audio tracks and identify abrupt
level changes, clipping, unwanted silence, and music that masks speech.
Use only the supplied music if music is appropriate.

Choose a loudness target for the intended destination and explain it.
Apply conservative processing, preserve the natural voice, and add clean
audio boundaries at cuts. Do not treat every pause as a mistake.

Export a short before-and-after sample and report the measurements available.
Tell me which issues need a listening comparison. Preserve the original
audio and save any settings used for the revised mix.

This is an instruction to assess and improve the track, not a promise that damaged audio can be fully restored. FFmpeg includes a loudness-normalization filter, but a measurement cannot decide whether aggressive cleanup makes a voice sound worse. Listen to the comparison. FFmpeg loudness filter.

6. Turn the approved edit into a vertical clip

A vertical export is a new composition problem. A widescreen screen recording may become unreadable if it is reduced to a narrow strip. A moving presenter may leave a static crop.

Decide what the viewer must see in each section. That may call for a crop around the presenter, a separate view of the interface, a stacked layout, or a background canvas that preserves the whole recording.

Prompt: create a vertical version

Create a vertical version from the approved edit and original source files.

Target: 1080x1920, 45–60 seconds.
Keep one complete idea and preserve any caveat needed to understand it.

Inspect the framing across the selected shots.
Use a crop only when the important subject remains visible.
For wide interface footage, use a readable layout that preserves the
control or result being discussed. Propose a stacked layout if useful.

Keep captions clear of the subject and the intended platform's interface.
Show sample frames from the opening, demonstration, and ending.
Export outputs/vertical-v1.mp4 and preserve the editable source.

The dimensions and duration here are an example brief. Check the destination’s current requirements before delivery.

A practical check is to watch the video on a phone at normal size. If the viewer must pause and zoom to read the demonstrated setting, the layout needs another decision. Enlarging a caption will not make the underlying interface legible.

7. Add captions after the wording and cuts are settled

Captions have two separate jobs: represent the speech accurately and remain readable in motion.

Ask for an editable subtitle file even if the final video will have captions burned into the image. It makes corrections and future versions easier.

Three deliverables are commonly confused:

Prompt: caption the approved cut

Create captions for the approved edit.

Use verified dialogue and timing. Check the spellings in brief.md.
Keep captions faithful to what was said; flag uncertain words for review.
Group words into readable phrases and avoid covering important visuals.

Generate:
- An editable SRT file timed to the final edited video.
- A captioned preview using the available rendering tool.
- A short list of uncertain words or timing problems.

Show representative frames so I can assess size, contrast, and placement.
If FFmpeg subtitle burn-in is unavailable, identify the missing capability
and use another already-available renderer if it can meet the brief.

Source captions must be retimed after cuts, reordering, or speed changes. An SRT aligned to the original ten-minute recording will not automatically align to a 60-second recut.

For an explainer, begin with calm, readable phrases. Word-by-word animation is an aesthetic choice, not a requirement for good captions.

8. Use Remotion for a reusable video format

Remotion is useful when the video is built from recurring elements: a title, presenter footage, a product screen, captions, and a closing card. The project can retain those elements as code that Codex can revise.

Remotion’s official documentation now describes a plugin for Codex in the ChatGPT desktop app. It includes agent skills and can help create a project. The documented installation route is to search for Remotion in the Plugins tab; its usage example invokes $remotion. OpenAI recommends starting a new chat after installing a plugin so the bundled capabilities become available. Remotion’s Codex plugin, OpenAI plugin setup.

Check the current Remotion system requirements and license for your environment and commercial use. Plugin availability does not remove the framework’s dependencies or license conditions.

Prompt: build a reusable Kingy.ai explainer

Use the installed Remotion plugin and skills to create an editable
30-second Kingy.ai explainer project.

Use only the supplied logo, screenshots, and approved text in brief.md.
Create:
- A short opening title.
- Three scenes, each explaining one point.
- A closing card with "Read the full guide at Kingy.ai."

Use a restrained design with readable text and enough time to understand
each screen. Keep the text editable and the assets organized.

Create a landscape composition at 1920x1080 and 30 fps.
Build a separate vertical layout if the material supports one.
Do not stretch the landscape layout into portrait.

Run the relevant project checks, open the preview, and save the project.
After I review the preview, render the approved composition to MP4.

Here, preview approval is a deliberate creative checkpoint for a new visual format. Once you like the template, later edits can proceed with narrower instructions.

For example:

Update the existing explainer template using the replacement screenshots
and approved copy in brief-v2.md.

Keep the established type, colors, scene structure, and motion.
Adjust line breaks and timing where the new copy requires it.
Render one short sample of the densest scene, then export the full update
if the sample passes the existing readability checks.

A framework project is editable, but it is not automatically a native Premiere Pro, Final Cut Pro, or DaVinci Resolve timeline. If you need an interchange file or an editor-specific project, specify that requirement and test an import before building the full workflow.

For readers who want to understand the export step, Remotion’s CLI takes a project entry point, composition ID, and output path. Ask Codex to report the actual command for your project, since those names vary. Remotion render reference.

9. Use HyperFrames for motion graphics and footage compositions

HyperFrames provides an agent-oriented route to making and revising video projects. Its current quickstart documents skill installation, an initial prompt, and a local preview. It also supports working with existing footage, so a project can combine a recording with titles or other graphic elements. HyperFrames quickstart, working with images and video.

If it is already installed, ask Codex to use the available HyperFrames skills. Otherwise, follow the current quickstart and install only what your chosen workflow needs. The documented manual skill-install command is:

npx skills add heygen-com/hyperframes

Run it in the intended project folder, choose the appropriate skills, and reopen the coding agent as the setup instructions direct. This downloads and installs third-party instructions; review the package and installation scope.

Prompt: add a focused graphic sequence

Use the installed HyperFrames skills to create a 12-second graphic
sequence for the supplied product demonstration.

Format: 1920x1080, 30 fps.
Assets: assets/logo.png and the approved screenshot in assets/.
Copy: use the exact three points in brief.md.

Structure:
- Introduce the subject.
- Explain one visible feature using the screenshot.
- End with "Full walkthrough at Kingy.ai."

Make each text block readable at normal playback speed.
Use purposeful motion and keep the screenshot legible.
Use local assets and retain an editable project.

Run the project's checks and open the preview.
After preview approval, render outputs/graphic-v1.mp4.

HyperFrames’ documentation provides npx hyperframes preview for the preview and npx hyperframes render --output … for export. It also distinguishes project checks from reviewing the finished file. Confirm available flags against the installed version. HyperFrames CLI, finishing and exporting.

There is relevant history on Kingy.ai: the site’s Codex and HyperFrames motion-graphics test records a rendered MP4 and a revision prompted by overlapping elements and an unwanted time marker. That earlier demonstration is useful context for the workflow; it does not establish Astra-specific performance.

For a precise visual revision, use:

In the current project, the headline and screenshot overlap during the
middle scene. Move the headline into a dedicated area above the screenshot,
reduce the copy to the approved shorter version, and preserve readable size.

Render a frame from the overlap point and a preview of that scene.
When the overlap is resolved, save the revised project and report the change.

10. Use an existing editor through Computer Use

If your work already lives in a desktop editor, Computer Use is another route worth testing.

OpenAI documents Computer Use for supported macOS and Windows environments in Codex and ChatGPT Work. Setup includes the Computer Use plugin, app access, and operating-system permissions. On macOS, Screen Recording and Accessibility permissions are required. Windows operation uses the foreground desktop. Computer Use setup.

This establishes the ability to operate graphical interfaces. It is not a certification that every Premiere Pro, Resolve, Final Cut, or CapCut operation will work reliably.

Start with a bounded task in an app you already use:

Use Computer Use in [editor name and version].

Work in a duplicate of [project name], saved as [new project name].
On the duplicate timeline, trim the opening to begin at [verified timecode]
and export a short review file using the existing approved export preset.

Inspect the current UI before acting. Verify the selected timeline and
output destination. Save the duplicate project and check that the export
exists and opens.

If the requested control is unavailable or ambiguous, describe the exact
obstacle and the step that needs my input.

Replace every bracketed field before sending it.

Choose a test that proves the operation you need. If your larger job involves caption styling, test caption styling. If it involves exporting a specific preset, test that export.

Where a dedicated integration or documented scripting interface is available, prefer it for repeatable operations. OpenAI makes the same distinction between structured integrations and graphical control in its Computer Use guidance.

11. Two small FFmpeg recipes you can verify

You can ask Codex to run these. Understanding the commands is useful when checking what it produced.

These examples assume a conventional SDR file with square pixels, readable video, and optional audio. They do not perform HDR-to-SDR conversion, repair damaged footage, or follow a moving subject. Run them from a project folder containing source/input.mp4 and an existing outputs/ directory.

The multiline commands use macOS/Linux shell syntax. In PowerShell, ask Codex to adapt them or use a single line.

Try it without recording your own footage

Download the example project. It contains an eight-second synthetic test-pattern video with a generated tone, the three-second cut, the vertical output, and the commands needed to reproduce them. No camera recording, stock footage, or external generation service is required.

Extract the archive and open its folder in Codex. The existing outputs let you compare the results immediately. To make fresh versions, use:

Use source/input.mp4 in this example project.
Inspect its metadata, then reproduce the two editing operations described
in README.md: a three-second cut beginning at two seconds, followed by a
1080x1920 centered vertical crop.

Save fresh files as outputs/my-cut.mp4 and outputs/my-vertical.mp4.
Use the existing outputs as references and preserve them.
Check the new files' duration, dimensions, video and audio codecs,
and complete-file decoding. Explain how much of the wide image the
vertical crop removes.

The moving pattern makes the crop visible, and the audio tone checks that an audio track survives the edit. This is a reproducible technical exercise. It does not demonstrate dialogue editing, storytelling, caption quality, or automatic subject tracking.

Inspect the input

ffprobe -v error -show_format -show_streams -of json source/input.mp4

This reports container and stream information. It is a technical inventory, not a review of the story or a transcript. FFprobe documentation.

Recipe A: export a three-second cut

ffmpeg -n -ss 00:00:02 -i source/input.mp4 -t 3 \
  -map 0:v:0 -map "0:a:0?" \
  -c:v libx264 -crf 20 -preset medium -pix_fmt yuv420p \
  -c:a aac -b:a 192k -movflags +faststart \
  outputs/test-cut.mp4

This starts at two seconds and creates a three-second output. The optional audio mapping allows a silent source. The command uses H.264 video and AAC audio; -n prevents overwriting an existing file.

The video is re-encoded. With FFmpeg’s default accurate seeking, transcoding allows material before the requested position to be discarded. A stream-copy trim has different boundary behavior and should not be presented as an interchangeable precision cut. FFmpeg seeking options.

The +faststart option moves MP4 metadata toward the beginning of the file for playback startup. FFmpeg MP4 format options.

Recipe B: make a centered vertical crop

ffmpeg -n -i outputs/test-cut.mp4 \
  -map 0:v:0 -map "0:a:0?" \
  -vf "scale=1080:1920:force_original_aspect_ratio=increase,crop=1080:1920,setsar=1" \
  -c:v libx264 -crf 20 -preset medium -pix_fmt yuv420p \
  -c:a copy -movflags +faststart \
  outputs/test-vertical.mp4

This fills a 1080×1920 canvas and crops the excess at the center. It is a mechanical crop. Inspect whether the subject survives it. Use the earlier vertical-edit prompt when the layout needs editorial judgment.

Upscaling a small source increases its pixel dimensions without restoring missing detail. The test fixture for these recipes was intentionally small; its enlarged output should not be treated as a quality example.

The preparation check produced two three-second files at 30 fps: a 640×360 cut and a 1080×1920 vertical version, both H.264 with AAC audio. Both decoded completely without reported errors. This verifies the demonstrated technical operations on that fixture, not lip sync, caption accuracy, or a human-footage editing result.

12. Review the actual exported file

A completed command is useful evidence. It does not tell you whether the edit communicates clearly.

For a short deliverable, watch the entire exported file. Check the opening, every cut, caption timing, the demonstration, and the ending. Listen to the audio rather than inferring its quality from the presence of an audio stream.

For agent-assisted checks, be explicit about what the tools can verify:

Check the actual exported file against brief.md.

Verify duration, dimensions, frame rate, video codec, audio presence,
and whether the complete file decodes without errors.
Inspect frames at the opening, ending, and significant cut points.
Check captions against the final timeline and flag uncertain wording.

Report separately:
- Technical checks performed and their results.
- Visual or audio checks performed with available tools.
- Anything still requiring human playback review.

Fix a material technical defect if found, then repeat the affected check.
Return the final file and editable project location.

Frame samples can reveal cropped text or overlaps. They cannot establish smooth motion or continuous lip sync. A model should not claim it “watched and listened to the whole video” if its tools only returned metadata and still images.

Keep feedback tied to the specific version you reviewed. A timestamp from v1 may refer to a different moment after v2 changes the opening.

13. Manage cost and rendering time

There are several possible costs:

OpenAI says task complexity, context, model choice, and tools affect usage. The current pricing documentation also says ChatGPT Work and Codex share usage. Use your account’s live usage information instead of assuming a fixed price per video. OpenAI usage and pricing.

HyperFrames says local rendering does not consume HeyGen credits; optional hosted and media services may have separate costs. Remotion’s license distinguishes eligible free use from uses requiring a company license. HyperFrames cost note, Remotion licensing.

A higher reasoning setting gives the model more room to work through a difficult problem. It does not directly make your video encoder faster.

The practical ways to contain a project are to approve the structure early, render short samples while settling a visual style, and reuse the project for revisions. Ask for full-resolution delivery once the sequence and layout are stable.

14. Fix common problems without restarting everything

If the problem is a missing dependency, fix that dependency or choose a supported route. Repeating the same prompt with more adjectives will not install a codec or create a transcription service.

15. Save the workflow once it works

After a successful edit, ask Codex to record the parts you want repeated. Keep it short enough to remain useful.

Prompt: preserve a reusable production recipe

Document the approved workflow in this project.

Include:
- Expected input files.
- Output formats and naming.
- The established caption and graphic styles.
- The actual preview and render commands.
- Required dependencies and versions.
- The checks that caught real problems in this project.

Create a concise reusable instruction file or skill if appropriate.
Keep footage, credentials, and project-specific claims out of the reusable
instructions. Do not install it globally or publish it.

Make the next run easy to start with replacement footage and an updated brief.

For a first attempt, one short video is enough: a few source files, one clear message, a saved project, and an MP4 you have watched. Keep the version that meets the brief. Reuse that project when you want the next clip.