AI Automating Video Edits
Markdown--- name: ai-automating-video-edits description: Analyzes viral short-form video styles (captions, zooms, SFX, pacing, color grade, brand elements) and recreates them in DaVinci Resolve using AI-assisted editing, learning and refining style from examples or instructions over time. Use when finding viral clips to repurpose, building an AI editing workflow for YouTube/Instagram clips, or training ChatGPT to replicate and evolve a specific editing style (the STEAL framework). ---
Give ChatGPT this prompt structure to start a new clip project:
STEAL FRAMEWORK - New Clip Project
NICHE: [your niche, e.g. podcast/business/motivation]
SOURCE: [link or description of viral clip to steal style from]
GOAL: Find the strongest moment, replicate the editing style, then adapt it to my brand
Tasks:
1. Research the source clip's style: caption style, zoom pattern, SFX usage,
pacing/cut frequency, color grade mood, on-screen brand elements
2. Transcribe my raw footage (I'll paste transcript below)
3. Identify the single strongest 30-90 sec moment for virality in my niche
4. Give me a full edit plan: cut points, caption timing/style, zoom moments,
SFX placement, music suggestion, color grade direction
5. Output it in a format I can execute in DaVinci Resolve (or feed to my AI
editing tool)
TRANSCRIPT:
[paste]
Progress:
- Step 1: Source — find a viral clip in your niche (ask AI to research trending formats)
- Step 2: Download + transcribe the clip (or your raw footage)
- Step 3: Style analysis — have AI break down the reference clip's editing DNA
- Step 4: Moment selection — AI finds the strongest segment from transcript
- Step 5: Color grade (manual, or instruct AI tool to apply a LUT/style)
- Step 6: AI-assisted edit pass — captions, cuts, zooms, SFX, music
- Step 7: Review — prompt refinements or manual tweaks
- Step 8: Feed results back to AI memory/project so it "steals" what worked
Step 1-2: Source & Transcribe
- Ask ChatGPT to research trending moments in your niche (e.g., podcast highlights, debate clips, emotional beats). Give it the niche + platform + current trends context.
- Download clip, transcribe using Whisper or Resolve's built-in transcription.
- Export transcript as a clean
.txtor.srt— this is what you feed ChatGPT, not the raw video.
Step 3: Style Analysis (the "Steal" part)
Prompt ChatGPT to deconstruct reference content into a reusable style profile:
Analyze this clip's editing style and output a STYLE PROFILE:
- Caption style: font, animation, capitalization, emphasis words, placement
- Zoom pattern: frequency, trigger points (punchlines, emphasis, silence breaks)
- SFX: types used (whoosh, impact, pop) and what triggers them
- Pacing: avg cut length, jump cut frequency
- Color grade: mood, contrast, saturation, warm/cool
- Brand/overlay elements: logos, lower thirds, intro/outro style
Save this as [CreatorName]_StyleProfile for reuse.
Store these style profiles in project files/custom GPT memory so they compound over time.
Step 4: Moment Selection
Feed the transcript and ask ChatGPT to flag the highest-potential segment using hook/payoff/retention criteria (strong opening line, emotional peak, controversial statement, punchline).
Step 5: Color Grade
Either do it manually in Resolve first, or instruct your AI editing tool to apply a grade matching the style profile's color notes before captions/SFX are added.
Step 6: AI Edit Pass
Use the style profile as a direct instruction set for your AI editing tool (e.g., auto-caption/SFX tools plugged into Resolve):
Apply [CreatorName]_StyleProfile to this edit:
- Captions: [specs from profile]
- Zooms: punch in 8-12% at emphasis words, punch out on new sentence
- SFX: whoosh on zoom-in, pop on text reveal, impact on punchline
- Music: [genre/mood] under -18dB, duck on speech
Step 7: Review & Refine
When something's off, prompt corrective instructions rather than redoing everything:
The zoom on "that's insane" felt late — trigger it 0.3s earlier.
Captions are too small for mobile — increase 20%.
Manually override anything faster to fix by hand than to prompt.
Step 8: Feedback Loop (Learning)
After each clip, tell ChatGPT what worked/didn't (retention data, comments, your own judgment):
This clip got 2x avg views. Update [CreatorName]_StyleProfile:
- Faster cuts in first 3 seconds worked well, keep under 1.5s avg
- SFX felt overused after 0:20, reduce frequency in back half
This is the "steal and improve" loop — each project session should end by updating the style profile file, not just the chat memory, so it persists across projects.
Example 1: Input: "Find a viral podcast moment in the self-improvement niche I can clip." Output: AI returns a specific clip (e.g., a MrBeast podcast moment) with a one-line reason it's viral (strong hook + payoff structure), plus a suggested caption/SFX style to match current trends in that niche.
Example 2: Input: Transcript of a 20-minute raw podcast pasted into chat with prompt "find the strongest 60-second moment for a short." Output: AI returns timestamped segment with the exact quote, explains why it's the strongest (tension + resolution), and suggests a hook line for the first 2 seconds of the clip.
Example 3: Input: "Apply the MrBeast_StyleProfile to this transcript-based edit." Output: Full edit plan — caption style (bold yellow, word-by-word reveal), zoom points at 3 flagged emphasis words, SFX list with timestamps, suggested music track mood, and color grade note (high contrast, warm highlights).
- Keep a style profile library. One file per creator/style you've reverse-engineered. Reuse and refine instead of starting from scratch each time.
- Separate transcript from video. ChatGPT works on text — always extract transcript/SRT before prompting for creative decisions.
- Prompt in small, specific corrections. "Make it better" is vague; "tighten the pacing in the first 5 seconds" gets results.
- Close the loop every time. End each editing session by updating the style profile with what worked — this is what makes the AI "get better" instead of resetting each time.
- Let AI handle decisions, you handle judgment calls. Use AI for pattern recognition (pacing, SFX timing) and your own eye for final taste/brand fit.
- Name and version style profiles (e.g.,
MrBeast_v3) so you can track what's improved.
- Don't feed raw video to ChatGPT — it can't analyze footage directly; always work from transcripts/descriptions.
- Don't skip the style-breakdown step and jump straight to "copy this style" — vague imitation produces generic results; explicit style profiles produce accurate ones.
- Don't let feedback live only in chat history — chats get lost/reset; persist learnings in a project file or custom GPT instructions.
- Don't over-automate SFX/zooms — too much AI-default styling makes every clip feel the same; always pass final review before publishing.
- Don't forget niche context — a style that works for comedy clips may kill retention on serious/educational content; always re-confirm niche + tone before applying a profile.