Edit videos with AI.
The complete guide.

The video below was fully edited with GPT-6 Astra. It is a finished example of AI video editing. Watch the first minute in particular; that is my favourite part.

Pay attention to the pacing, cuts and visuals. The chapters below walk you through setting up your own editing workflow, with the skills and prompts to get started.

Start with a 30 to 90 second recording. A short talking-head clip or screen recording is enough to learn the process. You can move to a longer YouTube video once the setup works and you like the style.

• video-rough-cut: transcribes, chooses takes, assembles the dialogue and records the cut decisions.
• video-animate: plans useful visuals and keeps the graphics editable.
• video-review: checks the export and applies specific corrections.
• Templates: a brief, visual plan, review record and handoff for the next session.

A skill is a folder of reusable instructions and supporting files. Keep the whole folder when you install it. The editor connection is what lets the agent actually work on your timeline.

The starting route in this kit uses Codex on a Mac and Palmier Pro. HyperFrames is optional for custom motion graphics. The kit is free; your agent account and any optional services have their own limits and costs.

1. Open Codex and sign in to your account. Choose GPT-6 Astra if it is available to you.
2. Check the current Mac requirements on Palmier's download page, then install and open the editor. The kit's documented route uses macOS 26 or later.
3. Create a folder called My First AI Edit and open it as your project in Codex.
4. Unzip the kit inside that project. Put a copy of your recording in media/. Keep your camera originals separately.

My First AI Edit/
  vibe-editing-kit/   The unzipped kit
  media/             Your recording and assets
  .agents/skills/    Installed editing skills
  analysis/          Transcripts and edit plans
  graphics/          Editable motion graphics
  exports/           Your rendered videos

Read the Vibe Editing Kit in this project. Create the working folders described in SETUP.md without replacing existing files.

Install video-rough-cut, video-animate and video-review from the kit into this project's .agents/skills/ folder. Preserve the complete folders, references and scripts. Report any name collision instead of overwriting it. Verify that all three SKILL.md files and their references are present.

Tell me what is ready and what is missing. Do not edit my recording yet.

If the skills do not appear, start a fresh task in the same project. You can ask for a skill by its full name: “Use the video-rough-cut skill installed in this project.”

Already on Windows, Resolve or Premiere? Read guide/other-editors.md in the download. Those editor connections are alternative routes, not validated setups for this kit.

Palmier exposes editor controls through MCP. That connection lets Codex inspect a timeline, make edits and export a video. Open Palmier and create an empty test project first.

In Codex, open Settings → MCPs and add a custom Streamable HTTP server named palmier-pro with this URL:

http://127.0.0.1:19789/mcp

If you use the Codex CLI, the equivalent command is:

codex mcp add palmier-pro --url http://127.0.0.1:19789/mcp

Keep Palmier open. These settings come from Palmier's connection instructions.

Check that it can actually read your project

Use Palmier MCP to list projects and open my empty test project. Read its timeline and tell me the actual project name, resolution, frame rate and clip count. Do not change another project. If more than one project could match, ask me which one.

You want the real project details back. Next, make a tiny export before handing it a full recording.

In this empty test project, import my short sample from media, place about five seconds on the timeline with its linked audio, and export a review MP4 to a new file.

Read back the timeline, wait for export to finish, and verify dimensions, frame rate, duration and audio presence. Keep the source unchanged. Tell me which checks passed and which still require playback.

Play the file yourself. Confirm that the picture and sound work and it starts and ends where expected. Solve any connection or export problem here, while the project is still small.

“Make this look good” leaves too much to guess. Tell the agent what the video is for, what needs to survive the edit, and what you want it to sound like. Include a glossary for names, products and unusual words.

Help me fill in templates/BRIEF.md for the recording in media. Ask me for any essential details you cannot infer:
- Who is watching and what should they understand?
- Is this a 9:16 reel or a 16:9 YouTube video?
- What ideas and closing invitation must stay?
- What is the spoken language, and which names need exact spelling?

Match the source frame rate. Keep natural speech and meaningful pauses. Remove failed attempts and unnecessary gaps. Use readable captions. Start with the Studio recipe. No background music. Use sound effects sparingly and only when they help a reveal.

Make the spoken rough cut first and let me approve it before adding visuals. Save the agreed choices in BRIEF.md.

Turn the example into your own brief

Revisit the Astra example at the start of this guide and write down the specific editing choices you like.

Give the agent those notes alongside your own footage. If it cannot inspect a YouTube link, supply reference screenshots or a clip you have permission to use. Use this prompt to turn the reference into a direction for your edit.

My reference is the opening minute of https://www.youtube.com/watch?v=hlzdg_tgrxY . First tell me whether you can actually inspect the video.

If you can, describe the observable editing choices: pacing, shot changes, captions, graphics and sound. Separate what you observed from suggestions. If you cannot inspect it, ask for reference frames or my notes instead of guessing.

Suggest which choices suit my recording. Propose a short sample before applying a new style to the whole video.

This is where the story takes shape. The agent should choose complete takes, remove restarts and keep the intended explanation. A shorter cut is useful only if it still makes sense.

Use video-rough-cut on my recording in media. Read BRIEF.md, verify the source and confirm the correct editor project.

Transcribe the recording with word timestamps. Check names against my glossary. Choose complete takes, remove failed starts and unnecessary gaps, and assemble the spoken edit. Keep the intended story and meaningful pauses.

Check each join against the source audio. Present the rough-cut video, editable timeline, cut plan and any uncertain decisions. Stop before adding graphics.

Start with the editor's local transcription. If it gets a section wrong, correct the language and spellings, then recheck that section. Word timestamps are estimates, so listen around the cut points.

Give feedback with output timestamps

Use the time in the version you are watching. A timestamp from the original recording can point to a completely different sentence after cuts.

In the current rough-cut export, at 00:12.4 the last word is clipped. Find the corresponding source audio and restore the complete word ending. At 00:24, keep the natural pause before the next sentence.

Preserve the rest of the approved edit. Export a new version and tell me which ranges changed.

The timestamps above are examples. Replace them with the moments in your own export.

This spoken cut is approved. Save its version and timeline in HANDOFF.md. Preserve its dialogue timing and audio for the visual pass.

The kit includes three original visual recipes. Pick one as your starting point, then adjust it to your footage. The illustrations below show the direction; the recipes are instructions and settings, not finished motion templates.

Use the Evidence recipe from video-animate for this project. Use my real screenshots as the proof. Create a short visual sample from the approved cut before styling the whole video.

Keep my face and important interface controls clear. Show me the sample, the caption placement and the graphics treatment. Preserve the approved dialogue timing.

Swap Evidence for Studio or Sketchbook. If something feels off, ask for one concrete change: smaller captions, a calmer entrance, more room around the screenshot or a longer hold.

Build visuals against the approved edit. For each beat, decide what the viewer needs to see. Some moments need a screenshot or diagram. Others work best with just you explaining the idea.

Use video-animate with my approved recipe and sample. Read the approved spoken cut and BRIEF.md. Create a visual plan against the current output timing.

Use my real screenshots and B-roll where they explain the point. Include clean presenter beats. Add readable captions timed to the edited dialogue, purposeful motion and only the sound effects allowed by the brief.

Keep demonstrations and my face clear. Preserve the approved dialogue and audio. Keep all graphics sources editable, render the visual edit and inspect the exported file.

Add HyperFrames only when you need it

Simple captions and editor keyframes can stay in Palmier. A custom diagram or designed motion sequence may need a separate composition. Use the official HyperFrames quickstart for that setup.

Set up HyperFrames for local rendering in this project's graphics folder using its official quickstart. Check the required Node.js version, FFmpeg and renderer browser. Install its core agent skills, record the version and prove the setup with a short local sample.

Keep paid generation and cloud rendering off unless I request them. Preserve my approved dialogue timeline. Keep the composition source and assets alongside every rendered insert.

Make revisions small enough to judge

In the current export, make the captions smaller and keep their horizontal alignment stable. Leave the opening explanation clean. At 00:18, show my actual screenshot instead of the text card and hold it long enough to read.

Do not change the spoken cut or the rest of the approved style. Export a new version and list the changes.

Replace the example time and visual request with your own feedback.

Review the exact file you will publish. A timeline preview can look fine while the exported MP4 has missing media, wrong dimensions or audio problems.

Use video-review on the latest exported file. Compare it with BRIEF.md, the cut plan and the visual plan.

Check speech boundaries, sync, caption spelling and timing, visual placement, missing media and actual export settings. Run a full decode check if the tools are available. Fix verified defects, preserve approved sections and re-check the revised file.

Give me the new export, editable project, issue list and a clear account of what you inspected or listened to. Tell me exactly which checks still need my playback.

Before you call it finished

• Watch the entire final file with sound, including the last few seconds.
• Listen for clipped words, unnatural joins and sound effects covering speech.
• Check lip sync and caption timing, especially after a revision.
• Read names and product labels. Make sure captions stay clear of important footage.
• Confirm the intended story and closing invitation are still there.
• Reopen the editable project and check that its media and graphics are available.

Ask for a handoff you can return to next week:

exports/video-v03.mp4
Editable editor project
analysis/cut-plan.json
analysis/visual-plan.md
analysis/review.md
graphics/ with composition sources and assets
HANDOFF.md with the current version and next steps

Technical checks can confirm that the file decodes or contains audio. You still need to listen to it and decide whether the edit works.

Once you approve an edit, save the choices that made it work. Your next video can start with the right caption placement, pacing and sound preferences already written down.

Turn this approved workflow into a reusable personal editing skill in this project. Include my corrections, pacing preferences and verification steps.

Capture the actual caption settings, visual treatments, motion timing and audio choices that worked. Separate general preferences from decisions specific to this video. Keep source paths, timecodes and source-to-output mapping in this video's handoff.

Do not replace the three kit skills. Create a separate personal-style skill and show me what you saved.

Useful rules sound like “hold the screen recording while the action happens” or “keep the pause before a punchline.” One-off notes such as “show this screenshot at 00:18” belong in that video's project.

Use my saved personal style with video-rough-cut, video-animate and video-review. Verify the new source and editor project first.

Reuse the approved visual and pacing choices, but plan the cuts and beats from this recording. Read the new brief, keep the source unchanged and let me review the spoken cut before adding visuals.

Keep the first error, app version and project name. Change one thing at a time so you can tell what fixed it.

Keep these references nearby

The download includes SETUP.md, FIRST-VIDEO.md, STYLE-PRESETS.md, SAVE-YOUR-WORKFLOW.md, SOURCES.md and VALIDATION.md. For current tool instructions, use Palmier's MCP docs, the official skills guide and HyperFrames' quickstart.