8 AI video editing prompts for a better first cut
Use AI as an edit-planning assistant, not as proof that the edit is finished. Paste a timestamped transcript or shot log, name one job, set the output format, and tell it what it must not invent. These eight prompts cover a rough cut, opening, captions, B-roll, pacing, transitions, color notes, and repurposed clips. Check every suggestion against the footage before changing the timeline.
01 · SOURCE BEFORE PROMPT
What should you give an AI before asking it to edit?
A useful editing prompt starts with evidence the model can point back to. OpenAI recommends clear, specific instructions with enough context, while Anthropic recommends stating the desired format and constraints. For an edit plan, that context is usually a timestamped transcript, a shot inventory, or both.
OpenAI: prompt engineering best practices for ChatGPTAnthropic: prompting best practices
| Input | What to include | What it prevents |
|---|---|---|
| Timestamped transcript | Exact words, speaker changes, and uncertain text. | Invented quotes and vague cut points. |
| Shot log | File names, time ranges, framing, and usable moments. | Suggestions for footage that was never recorded. |
| Editing job | One goal such as rough cut, captions, or B-roll. | A generic answer that mixes unrelated decisions. |
| Output schema | Named columns, labels, or a file-shaped draft. | Advice that cannot be transferred to the timeline. |
| Guardrails | Facts, claims, footage, rights, and context to preserve. | A cleaner edit that changes what actually happened. |
Scroll horizontally to read all columns.
02 · CHECK THE WORKFLOW
What AI editing can and can’t do yet (October 2026)
A motion graphic, a silence-removal pass, and a finished video are different jobs. Before copying a prompt from a tutorial, check what went into the edit and what someone still had to fix. These demonstrations show useful pieces of the workflow, not a guarantee that your footage will come out ready to publish.
| What the demonstration shows | What to check before relying on it |
|---|---|
| Reference-led motion graphics. RandomAI shows a reference video and detailed instructions used to generate Remotion animation code, then shows the output. RandomAI: reference-led motion graphics demonstration Creator demo: a JavaScript alternative. Futurepedia shows Claude drawing explainer animations frame by frame with plain JavaScript rather than requiring Remotion or HyperFrames. Its style prompts include example source code. Choose a style and topic, review the storyboard, then approve generation. Futurepedia: JavaScript explainer animation (creator demo) | The working environment was already configured. Check the animation’s text, branding, layout, and timing; generating a graphic is not editing your raw footage into a complete story. Futurepedia also links a resource advertised as 86 styles. Its Style Atlas signup page asks for an email and other details, with a notice about marketing contact from HubSpot and Futurepedia. It is not an ungated prompt download. |
| Editable changes inside After Effects. Shadow shows a connected workflow that creates layers, text, and keyframes through repeated instructions. Shadow: connected After Effects workflow | Confirm your editor connection works and the changes remain editable. Watch the revisions as well as the final preview; this is a multi-step workflow. |
| Creator demo: an editable animation draft. Vane Motion makes a bouncing ball by hand, then asks a Codex session labeled GPT-6 Astra High to build a version in After Effects. The video shows the composition settings, ball and ground layers, and editable keyframes. Vane Motion: manual and AI-assisted animation (creator demo; course and membership promotion) | The creator compares the versions and finds that easing and motion details still need attention. Inspect the squash and stretch, speed curves, and loop seam before accepting an animation. This tutorial includes course and membership promotions. |
| Creator demo: a less decorated talking-head edit. mino.mp4 describes the instructions he gives Claude for captions, visuals, and pacing. The video includes comment-to-get offers for his editing skill and MCP. It shows examples of the look he wants, rather than a complete editing session. mino.mp4: caption and pacing preferences (creator demo) | He suggests a plain font with two to three words on screen at a time, and relevant screenshots instead of decorative emoji or icons. Check that each caption group is readable and the screenshots are yours to use. He says he removes pauses longer than 0.1 seconds. That is his aggressive pacing choice, not a general editing standard. Test a short section and keep pauses needed for breathing, emphasis, or a demonstration. He also asks Claude to identify redundant lines and sentences unrelated to the hook. Review those suggestions before cutting, so necessary context and qualifications stay attached to the claim. He describes connecting a Google Drive B-roll folder, but does not show the full connection or authorization process. Confirm your own setup can access the intended folder and that you have permission to use its assets. |
| Rule-based silence removal. At about 2:57, a creator shows DaVinci Resolve’s Clip → Audio Operations → Ripple Delete Silence command and its Remove Silence settings. Watch the silence-removal demonstration | This command handles one cleanup task. Test a short section with settings suited to your recording, then listen for clipped words and intentional pauses. The creator removes mistakes separately. |
| Creator demo: prepare a long-form style library. Joseph shows Claude Code and Remotion working from 10 of his own reference videos. He organizes their design patterns into an editing skill, plans scenes, and reviews design sheets and an intro render before expanding the work in sections. He describes the resulting edit as an 18-minute video. Joseph: reference-led long-form workflow (creator demo; course promotion) | Ten references are this creator’s input, not a required minimum. He explains part of the body-editing process after losing its screen recording. The tutorial promotes his course; he says this particular long-form skill was not yet available to students at the time. Expect to supply your own references and review each section. |
| Captions from an existing transcript. Descript’s demo shows captions being added by selecting a preset, with the captioned result visible immediately afterward (00:11 to 00:15). Descript: add captions from a preset (00:11 to 00:15) | The transcript is already in the project. This shows a caption preset being applied, not a prompt-driven edit or the full import-to-transcription process. Review the words and timing against your audio. |
| A result to compare with a human edit. A creator compares AI, a hired editor, and his own version using the same footage. Watch the three-version editing comparison | This is one creator’s comparison, not a general quality ranking. He asks for music and sound effects after the AI’s first pass. Allow time for revision rather than expecting ready-to-publish results from one prompt. |
Scroll horizontally to read all columns.
For talking-head footage, the reference is only the start. You also need a working local setup and a way to review the edit before rendering the whole video.
Creator demo: review a storyboard before rendering
Brand Nat’s walkthrough shows a local Claude Code workflow with a storyboard and a visual style menu. Her five steps include tool preparation, which a prompt alone cannot replace.
- Choose a local project folder. Open your working folder in Claude Code so the session can access the media you intend to use.
- Prepare the editing tools. The creator asks Claude Code to download the tools on her list: FFmpeg for cutting and mixing, HyperFrames for captions and graphics, and yt-dlp for retrieving the reference. The video does not supply installation commands or versions. Check what will be installed, and use only media you have permission to access.
- Add your raw talking-head file and a reference. She shows a local video attachment and an Instagram link, and also suggests TikTok references. Explain which editing techniques you want to borrow; first check that the session can actually read your file and the reference.
- Ask for a storyboard first. Review its scene order, captions, and layouts. Request changes there before approving a full render. The creator suggests this to avoid spending time and tokens on a direction you do not want; the walkthrough does not measure the savings.
- Save the approach once you are satisfied. Ask Claude Code to turn the accepted workflow into a reusable skill and a visual style menu for captions and layouts. The video shows the menu, but not the skill file’s location or a complete demonstration of loading it in another project.
Brand Nat: local talking-head workflow (creator demo)
Creator demo: save your style and approve the cut list
If you keep repeating the same caption instructions, save them with the project. kienobifilms recommends saving reusable editing instructions in a CLAUDE.md file inside a dedicated project folder. Include the font, colors, caption placement, and areas covered by Instagram’s interface. Describe those preferences once, then reuse the file for later edits.
For the rough cut, the creator asks Claude Code to remove pauses and retakes from the raw clip, keeping the last take. Ask it to list the proposed cuts before changing the footage. Review that list yourself before approving the edit, especially where a pause or repeated line is intentional.
For a reference reel, his workflow uses an uploaded video file, sampled frames, and a transcript to build a style guide. Use a file you have permission to use, rather than assuming a pasted link gives the assistant access to the full video.
kienobifilms: reusable instructions and a proposed cut list (creator demo)
A polished demo can leave out setup and failed attempts. Vlad Sateev discusses those omissions and his own difficulties with editing workflows. Treat that as a creator’s experience, not a controlled test of every product. Before committing your full project, try one short sequence and check that you can repeat the steps with your own files.
Vlad Sateev: what editing demonstrations leave out
Automatic captions still need a correction pass. YouTube recommends reviewing them against the audio. Use the caption prompt below to format a checked transcript and flag uncertain words or timings.
03 · COPY THE JOB, NOT THE HYPE
8 AI video editing prompts to copy
Replace every bracketed field. Keep file names and timestamps intact. Each prompt asks for a plan or draft, never a claim that the AI touched your media. If your assistant can inspect images or video, give it only the assets you have permission to share.
You are planning a rough cut, not editing the media files. INPUTS - Target viewer: [who] - Target runtime: [length] - Required facts or moments: [list] - Timestamped transcript: [paste] - Shot log with file names and time ranges: [paste] TASK Build a rough-cut decision table with these columns: 1. source time range 2. KEEP, CUT, MOVE, or VERIFY 3. exact spoken line or shot description 4. proposed timeline position 5. reason tied to clarity, continuity, or the stated goal 6. editor action RULES - Use only words and shots present in the inputs. - Never invent a quote, reaction, camera angle, result, or piece of B-roll. - Mark uncertain transcript text or missing visual evidence as VERIFY. - Preserve required facts even when they slow the pace. - Do not claim that any cut has been applied or exported. Finish with the proposed runtime and a short list of unresolved checks.
Give it: A timestamped transcript, shot log, target length, and must-keep facts. Get back: A source-to-timeline table with keep, cut, move, and verify decisions. Why this works: This makes the model show where every recommendation came from. You can reject one row without rebuilding the whole plan.
Act as an opening editor. I need three truthful options for the first 8 seconds. INPUTS - Intended viewer: [who] - Verified payoff in this video: [what the footage really shows] - Timestamped transcript: [paste] - Shot log: [paste] FOR EACH OPTION, RETURN - exact in and out source timestamps - verbatim spoken words, in order - the first visible shot - the question or promise created - what later moment pays it off - any continuity risk RULES - Do not rewrite spoken words and present them as recorded audio. - If a new line would help, label it PICKUP TO RECORD and keep it separate. - Do not imply a result, test, or personal experience missing from the footage. - Do not choose a visual that is absent from the shot log. - Do not claim one opening will increase retention. Rank the options by clarity, not predicted virality.
Give it: The transcript, shot log, verified payoff, and intended viewer. Get back: Three openings with exact source references and any pickup line labeled. Why this works: A stronger opening is useful only if the footage can deliver it. Source references stop a catchy rewrite from becoming a false promise.
Turn the checked transcript below into an SRT-shaped caption draft. INPUTS - Language: [language] - Speaker names and spelling glossary: [paste] - Reading constraints: [maximum lines and characters, if known] - Checked transcript with timestamps: [paste] TASK - Split captions at natural phrase boundaries. - Keep the original meaning and wording. - Use sequential SRT numbers and HH:MM:SS,mmm timestamps. - Put uncertain words in [[double brackets]]. - After the draft, list every name, number, technical term, overlap, and noisy section that needs review. RULES - Do not invent missing speech or translate unless I explicitly ask. - Do not paraphrase a claim to make it shorter without marking the change. - Do not guess speaker identity. - Do not claim the draft is ready to import. Return only the SRT-shaped draft and the verification list.
Give it: A checked transcript, names glossary, language, and caption rules. Get back: An SRT-shaped draft plus a list of words and timestamps to verify. Why this works: Caption formatting is mechanical, but names, accents, and timing still need a human pass before import or publication.
Create a B-roll plan for this edit using the footage inventory as the source of truth. INPUTS - Timestamped transcript: [paste] - Available footage and stills, with file names and time ranges: [paste] - Rights or disclosure notes: [paste] - Visual style constraints: [paste] RETURN A TABLE WITH 1. timeline moment 2. spoken idea the visual must support 3. available asset and exact source range 4. framing or crop note 5. duration estimate 6. status: AVAILABLE, VERIFY RIGHTS, or NEW SHOT NEEDED RULES - Never invent an available asset. - Use an asset only when its description supports the spoken idea. - Label every missing visual as NEW SHOT NEEDED; describe the shot, but do not claim it exists. - Keep licensed, sponsored, or AI-generated material visibly flagged for review. - Do not claim that any media was placed on the timeline. End with the three highest-priority coverage gaps.
Give it: The transcript, available asset inventory, and rights notes. Get back: A B-roll plan that separates usable footage from new shots to capture. Why this works: The prompt treats missing coverage as a production decision instead of silently inventing an asset that the editor cannot find.
Review this cut plan for pacing risks. Treat every conclusion as an editing hypothesis. INPUTS - Video goal and intended viewer: [paste] - Timestamped transcript: [paste] - Current cut or edit decision list: [paste] - Known shot, text, or topic changes: [paste] RETURN A TABLE WITH 1. timeline interval 2. observable issue in the supplied material 3. why it may slow comprehension or create overload 4. one edit to test 5. information or continuity that edit could damage 6. verification needed RULES - Do not infer audience retention, views, or algorithm response from the transcript. - Do not mark a pause as dead air when it serves a demonstration, emotion, or comprehension. - Do not remove required context to make the cut faster. - Use only the supplied timing and shot-change data. - Do not claim the edit has been made. Finish with one conservative pacing pass and one more aggressive pass.
Give it: A timestamped transcript, current cut notes, and known shot changes. Get back: Intervals to inspect, one proposed action each, and a stated tradeoff. Why this works: Pacing is not the same as constant speed. The output protects necessary context and keeps every diagnosis as a hypothesis until data or review supports it.
Create a transition plan for the proposed cut below. INPUTS - Ordered rough-cut table: [paste] - Audio notes, room tone, and music limits: [paste] - Any intentional changes in time, place, speaker, or topic: [paste] FOR EACH CUT, RETURN - outgoing and incoming source ranges - relationship: CONTINUOUS, NEW TIME, NEW PLACE, NEW TOPIC, or CONTRAST - recommended transition: HARD CUT, J-CUT, L-CUT, DISSOLVE, or HOLD - why the transition clarifies that relationship - audio or continuity check RULES - Default to a hard cut when no other transition has a clear job. - Do not add sound, room tone, music, or footage that is not listed. - Do not hide a continuity error with a transition; flag it. - Do not use a dissolve to imply time passing unless the story supports that meaning. - Do not claim the transitions were applied. End with the cuts that require an editor's judgment in playback.
Give it: The rough-cut order, audio continuity, and relationship between adjacent shots. Get back: A transition map with a default hard cut and exceptions justified. Why this works: Transitions should communicate time, place, continuity, or contrast. Starting with a hard cut avoids decorating every edit by default.
Act as a color-review assistant. Write correction notes, not a finished grade. INPUTS - Reference frames with shot IDs: [attach or describe each visible frame] - Camera, profile, and lighting notes if known: [paste] - Intended visual reference: [attach or link] - Delivery platform and color space if known: [paste] RETURN FOR EACH SHOT - visible mismatch relative to the supplied reference - correction goal for exposure, white balance, contrast, and saturation - skin-tone or product-color check when relevant - what must be measured on scopes - status: OBSERVED IN FRAME, PROVIDED METADATA, or ASSUMPTION TO VERIFY RULES - Do not invent camera settings, color temperature, exposure values, LUTs, or scope readings. - If no frame is supplied, say that visual assessment is not possible. - Separate technical normalization from creative look choices. - Preserve deliberate lighting differences unless I identify them as errors. - Do not claim that a grade has been applied or exported. Finish with a shot-matching order that starts from the best reference frame.
Give it: Reference frames, shot IDs, camera or lighting notes, and delivery target. Get back: Correction goals per shot, with measurements or assumptions flagged. Why this works: Text alone cannot reveal image data. Requiring frames and labeling assumptions keeps the plan useful without inventing exposure or color values.
Find up to three self-contained short clips in this longer video. INPUTS - Timestamped transcript: [paste] - Shot log: [paste] - Target platform and maximum duration: [paste] - Facts, disclosures, or context that must stay attached: [paste] FOR EACH CLIP, RETURN - exact source in and out timestamps - verbatim first line and final line - the single question or takeaway it can answer alone - required context inside the clip - matching available visuals - estimated source duration - any PICKUP TO RECORD, clearly separated from existing speech RULES - Keep the speaker's meaning intact. - Do not combine nonadjacent words into a sentence that was never spoken. - Do not remove a qualification, disclosure, or condition that changes the claim. - Do not invent B-roll, performance results, or a title presented as proven. - Do not claim the clips were cut or published. Reject any candidate that cannot stand alone within the stated duration.
Give it: A timestamped transcript, shot log, platform limit, and required context. Get back: Three clip candidates with exact boundaries and any pickup need labeled. Why this works: The model must find complete ideas already recorded, not manufacture a hook that changes the speaker's claim.
04 · MOVE FROM PLAN TO TIMELINE
How do you turn an AI answer into a real edit?
Treat the answer as an edit decision list. It can reduce the blank-page work, but the footage, playback, and final timeline remain the authority. Use this five-step pass to review the plan before applying it.
- 01
Trace every note to a source
Open the named file and source range. Reject any quote, shot, or timing decision that cannot be found.
- 02
Test on a duplicate sequence or a short section
Keep the original sequence. Try the highest-impact cuts on a duplicate before changing the full edit.
- 03
Watch for continuity and changed meaning
Check the audio, eyeline, action, disclosures, and every qualification that makes a claim accurate.
- 04
Validate captions and generated assets separately
Review spelling and timing before import. Confirm rights and disclosure needs before using generated footage.
- 05
Compare the published result with real evidence
Use playback review, comments, and platform analytics to test the hypothesis. Do not backfill a causal story from one view count.
If you are deciding whether this workflow is worth the time and subscriptions, use Kabo's guide to AI video editing cost. If a platform has already assembled the first cut, the checklist for reviewing Instagram First Draft shows the same source-first discipline in a product workflow.
05 · MATCH THE TOOL TO THE JOB
Which AI tool should you use?
A text assistant can help draft an edit plan from a transcript. For visual review, choose a multimodal assistant that can safely inspect the frames or footage you provide and point to what it reviewed. Creating a new clip calls for a video generator; changing the timeline and exporting calls for an editor.
Google Flow, for example, can generate clips from text, frames, or other ingredients. That does not mean a clip has been reviewed, licensed for your use, placed into your existing sequence, or published. Keep generation and editing as separate checklist items.
Google Flow Help: create videos in Flow
| Tool type | Useful for | Still needs verification |
|---|---|---|
| Text assistant | Transcript, caption, and decision-list drafts. | Source accuracy, timing, visuals, and final playback. |
| Multimodal assistant | Reviewing supplied frames or clips with text context. | What was actually inspected and what remains unseen. |
| Video generator | Creating a candidate shot when generation is the job. | Rights, disclosure, continuity, quality, and placement. |
| Video editor | Timeline changes, playback, captions, and export. | Meaning, claims, rights, and audience evidence. |
Scroll horizontally to read all columns.
06 · KEEP THE CLAIM SMALL
What can these prompts not verify?
The prompts cannot see footage you did not provide. They cannot prove that an edit caused retention, views, recommendations, or sales. They cannot establish rights to a clip, identify a person, confirm a disclosure, or know whether a line was recorded truthfully. They also cannot promise that an SRT-shaped answer will import without errors.
YouTube's audience-retention report describes moments where viewers start or stop watching, but the report still needs context and is not instant. YouTube says the data typically takes one to two days to process. Use dips and spikes to decide what to inspect, then compare the video, traffic, comments, and your edit notes before forming a hypothesis.
07 · FAQ
Common questions about AI video editing prompts
Can ChatGPT or Claude edit my video from a prompt?
A chat assistant can plan edits from material you provide. Whether it can inspect media or change a timeline depends on the product and tools connected to it. Keep the prompt's output labeled as a plan until you verify a real timeline.
What if my transcript has no timestamps?
Ask for structural notes by paragraph or sentence, not precise cut points. Add timestamps before requesting an edit decision list. Automatic captions can be a starting point, but correct uncertain words first.
Can I ask AI to make an SRT file?
You can request an SRT-shaped draft. YouTube supports basic SRT files as plain UTF-8 text, but you still need to validate the numbering, timestamps, text, and import in your editor or platform.
Do these prompts work for long and short videos?
Yes, if you replace the runtime, platform, and context constraints. Do not force a long explanation into a short clip by removing a qualification that changes the claim.
08 · SOURCE LEDGER
Sources and method
Product and platform guidance was checked on September 16, 2026; the Descript caption demonstration was checked on October 20, 2026. The official sources support prompting principles, video generation boundaries, caption review, SRT formatting, and retention-report behavior. The eight prompts, input table, and five-step verification pass are Kabo editorial material. They are not product guarantees, tested performance results, or a claim that an AI system completed an edit.
- Descript: add captions from a preset (00:11 to 00:15)
- OpenAI: prompt engineering best practices for ChatGPT
- Anthropic: prompting best practices
- Google Flow Help: create videos in Flow
- YouTube Help: use automatic captioning
- YouTube Help: supported subtitle and caption file formats
- YouTube Help: measure key moments for audience retention
October 2026 creator demonstrations
These videos were checked October 20, 2026. They show the creators’ own setups and results; Kabo did not independently repeat their edits. The three-version comparison reflects one creator’s judgment, not an audience test. Setup, additional instructions, and human corrections still matter when you try the workflow yourself.
- RandomAI: reference-led motion graphics demonstration
- Shadow: connected After Effects workflow
- vidIQ: three edits from the same footage
- Vlad Sateev: what editing demonstrations leave out
Additional creator demos checked October 21, 2026
These tutorials show ways to plan animations, organize editing references, and review drafts. Your project may need different tools or setup steps. Futurepedia’s linked Style Atlas requires signup; review its form before sharing your details.
- Futurepedia: JavaScript explainer animation (creator demo)
- Brand Nat: local talking-head workflow (creator demo)
- Joseph: reference-led long-form workflow (creator demo; course promotion)
- Vane Motion: manual and AI-assisted animation (creator demo; course and membership promotion)
Creator demo checked October 23, 2026
Use this walkthrough for reusable style instructions and reviewing a proposed cut list before editing.
Creator demo checked October 26, 2026
mino.mp4 describes his caption and pacing preferences and shows examples. Kabo did not independently repeat the edit or measure its effect on editing time or results. The video includes comment-to-get promotions; the linked resources were not reviewed.
Draft-review creator demo checked October 27, 2026
This tutorial describes the creator's graphic density, draft review, and render scheduling. The storyboard above is illustrative; production results and hardware limits have not been independently reproduced.