The last step before a paid ad goes live is the one most often skipped: actually watching the finished file, all the way through, looking for the misspelled overlay and the caption that runs off the bottom of the safe area. analyze_video does that pass on demand — it extracts sample frames (uploaded as reusable JPEG URLs), describes the overall style and format, gives per-timestamp beats, reads the on-screen text, and optionally returns a speech transcript.
review_generation is the narrower, stricter sibling: it self-QCs a generated image or video against its ORIGINAL prompt via vision and returns a structured verdict — matches_prompt, a list of issues, a severity of ok, minor or broken, and a suggested_fix. Its own description says to offer it after a long job or when the user seems unhappy, and never to auto-run it on every generation, because it costs credits and latency.
Powered by
Deep, on-demand video understanding: extract sample frames (uploaded as reusable JPEG URLs), describe overall style/format, per-timestamp beats, on-screen text, and optionally a speech transcript. Works with ANY chat provider (provider-agnostic). Use for thorough analysis or when you need frames as reference images for generate_image_from_image / workflow assets / recreation — deeper than the automatic attachment preprocessing that only injects a short text note. Prefer this over guessing from a video URL alone.
What to tell the agent
Versely's agent maps this plain-English request directly onto analyze_video. You don't need to know the parameter names — just describe what you want.
“Watch this finished ad back before I publish it — check the on-screen text, the pacing, and whether it matches the brief.”
How it works
1. Run the deep read
analyze_video takes the video_url plus depth and focus, so you can point it at what you actually care about — 'check every on-screen caption', 'describe the pacing beat by beat' — rather than getting a generic summary.
2. Get the transcript if the words matter
include_transcript adds the spoken content, which is the fastest way to catch a claim in the voiceover that legal will not sign off on.
3. Grade a generation against its brief
For an AI-generated clip, review_generation compares the output to the original prompt and returns matches_prompt, issues, a severity (ok / minor / broken) and a suggested_fix — a decision, not just a description.
4. Fix and re-check
Act on the verdict with the relevant editing tool, then re-run the check on the corrected file rather than assuming the fix landed.
What it costs
Both reviews cost credits and time — review_generation's own guidance is explicitly not to run it on every generation. Reserve them for the files that are about to be published or paid to promote, where a broken caption is more expensive than the check. Ask the agent for estimate_cost before running a deep pass on a long video.
The formula behind that number: What does a finished 30-second AI ad cost end to end? — sums four unrelated meters, so no single rate predicts the total.
Limits & things to know
- This is a review, not a repair — it tells you what's wrong and suggests a fix, but the correction is a separate edit.
- review_generation needs the ORIGINAL prompt to grade against, so it's for AI generations rather than arbitrary uploaded footage. For that, analyze_video is the right tool.
- Don't auto-run these on everything. The tool's own instruction is to offer the check after long jobs or when something looks off, because each pass costs credits and latency.
- analyze_video samples frames rather than inspecting every one, so a single-frame glitch can still slip through a shallow pass — raise depth if that's what you're hunting.
Who uses this
- Final QC on a paid ad before the spend starts
- Catching a typo in a burned-in caption while it's still fixable
- Checking a client deliverable against the approved brief
- Pulling reference frames out of a finished cut for the next version
Frequently asked questions
What does the review actually return?+
analyze_video returns sampled frames as reusable JPEG URLs, a style and format description, per-timestamp beats, the on-screen text and optionally a speech transcript. review_generation returns a structured verdict instead: matches_prompt, issues, a severity of ok, minor or broken, and a suggested_fix.
Should I run this on every video I make?+
No — review_generation's own guidance is to offer it after long jobs or when something seems off, and never to auto-run it, because each pass costs credits and latency. Save it for the files that are about to be published or promoted.
Can it check a video I uploaded rather than generated?+
analyze_video works on any video URL. review_generation is the one that needs an original prompt to grade the output against, so it's specific to AI generations.
Does it fix the problems it finds?+
No. It reports issues and suggests a fix; applying it is a separate edit with whichever tool the problem calls for.
Related Versely tools
Related editing jobs
Check a Video Before You Publish It inside Versely
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.