This AI actually watches the footage and returns a visual QA report in seconds, saving time and reducing embarrassing mistakes.
Below is an AI-assisted summary of the key points and ideas covered in the video. For more detail, make sure to check out the full time stamped video above!
Why “AI that can see” is a big deal for video quality control
Most review workflows rely on someone manually watching the final cut and hoping everything is caught. That works until a typo goes live or a phone number or other confidential info appears on screen.
This approach tackles the usual pain points fast:
- Catches on-screen typos and spelling mistakes (not just spoken words)
- Flags sections that drag on too long, which helps retention
- Spots repeated content missed during a quick review
- Checks for technical glitches that break flow
- Scans for confidential information such as phone numbers or credit card details
It’s a sanity check at the end—done is better than perfect, but “done” shouldn’t mean “dodgy.”
The key setup: Descript + Claude Opus 4.5
The workflow runs inside Descript. The crucial detail: the only model referenced that actually “looks inside” the footage is Claude Opus 4.5.
Why that matters:
- Other options might summarize or respond based on text only
- Claude Opus 4.5 returns details that prove it’s using the visuals, like shirt color, desk items, or a smartwatch
Token usage is higher, but it’s the obvious choice when visual QA is the goal.
How to run an AI video review in Descript (fast, repeatable workflow)
Follow this process as the final step before publishing. It’s built to save time and give peace of mind.
- Export a low-quality version of the finished edit. Lower quality speeds things up. It doesn’t need to be 4K or even 1080p.
- Create a new project in Descript and import the video file. Any editing software can create the edit—just bring the final export into Descript.
- Drag the file into the sequence/timeline so Descript transcribes it. When transcription finishes, the preview and the fully edited cut appear.
- Open Descript AI and go to the instructions area (shown under Underlord). That’s where the AI is told exactly what to review.
- Select the model: Claude Opus 4.5. This setting enables the “actually watches the footage” behavior.
- (Optional) Run a quick visual-proof test question. Ask something like: “What colour is the T-shirt being worn in the footage?” The response should include visual details.
- Paste in a dedicated “video reviewer prompt” and run it. The prompt used runs through six phases of review, looking for mistakes, repeats, engagement opportunities, and confidential info checks.
- Copy the AI output into Google Docs or a Word doc for easier reading. The report can be hard to read inside the interface but becomes easy to scan once pasted into a document.
This turns quality control into a simple system: import → prompt → report → fix.
What to tell the AI to look for (prompt checklist)
Even without the full prompt text, the transcript spells out the big rocks to include in instructions:
- Typos and spelling mistakes, especially on-screen text
- Repeated content or sections that say the same thing twice
- Areas where the presenter talks too long, useful for tightening pacing
- Confidential information, including credit card details, phone numbers, and emails
The real win is that it reviews what viewers actually see—titles, overlays, and any accidental on-screen leaks.
How to read the “Video Quality Control Report”
Once the report is pasted into a doc, it’s structured so action can be taken quickly instead of wading through a big block of text.
1) Retention killers (where people might drop off)
This section calls out parts that could hurt engagement, such as:
- An extended talking-head intro that delays value
- Too many rhetorical questions before getting to the point
- Over-explaining a concept (example mentioned: “run and gun”)
- Suggested trims and cuts to tighten the flow
If the goal is adding value earlier, this is low-hanging fruit.
2) Clarity boosters (visual aids that improve understanding)
This section highlights what’s working and what could be clearer:
- Positive callouts when a visual demo supports the explanation, like showing finger movement on a phone
- Suggestions to add text callouts for specifics, such as listing formats like ProRes log and 10 bit
It’s basically a checklist for “Would the viewer instantly get this?”
3) On-screen text errors (typos, formatting, consistency)
This area shows where the workflow shines. The report can flag issues like:
- Formatting inconsistency (example: how “run and gun” appears on screen)
It may over-report minor things. That’s intentional—better to flag and confirm than to miss something unprofessional.
4) Audio/visual mismatches and technical glitches
The report checks for issues that can slip through late in the edit, including:
- Audio/visual mismatches
- Extended blank screens
- Extended silence
A clean result here is the quiet reassurance needed before something goes live.
5) Confidential information scan (the stress reducer)
This category checks for sensitive details that cause real anxiety:
- Credit card numbers
- Phone numbers
- Emails
Finding none is a major peace-of-mind win.
6) Quick summary (the TL;DR action list)
At the end, a short hit list of what matters most appears, such as:
- Tighten the intro
- Fix the on-screen “run and gun” formatting if needed
- Consider trimming a promotional section
That’s a clear win: faster review, fewer mistakes, and a tighter final product.
How to use the report without blindly following it
The workflow works best when the AI report is treated like a sharp assistant, not the final decision-maker.
Use it this way:
- Jump to flagged timestamps and watch the exact moment in Descript
- Decide if it’s a real issue or a harmless style choice
- Fix high-impact items first (confidential info, obvious typos, retention-killing rambles)
- Leave minor preferences alone if they don’t move the needle
Small, consistent improvements add up across every upload.
Publish with confidence
Run this AI review as the final pre-publish step to catch easy-to-miss items like typos, repeats, and confidential info. Keep the export low quality to speed processing, use Claude Opus 4.5 as the model, and paste the report into a document so it’s readable. Then fix what matters and ship with confidence.
