A harmless title can sit above footage that crosses the brand-safety line.
Content production · quality screening · human approval
Editorial automation system
Produces the day's video and image-and-text assets, then stops for editorial approval.
Running dailyOperational review
Signal → cause → change
TITLE ≠ FOOTAGE- TITLE ≠ FOOTAGE
Signal Root cause The deterministic gate reads titles. It never watches the video.
Change The publish gate stays with a person who must watch the finished asset.
What I built
I connected sourcing, screening, editing, and copywriting into one workflow. It delivers finished options each morning, remembers why work was rejected, and leaves the final publish decision to a person.
Complete workflow
the review of 2026-08-18
Yesterday's feedback becomes memory2026-08-18
The first real review
23in the calendar
1published
8sent back
All eight were comedy; the news line lost nothing. So the hard part is picking, not producing.
Review reasonsStored as memory
Gather today's candidate poolReddit · Twitch
Zendaya Acting in a Nutshellcomedy
A hippo waits for a watermeloncomedy
Ants carrying a buttoncomedy
YouTube doubles the watch hourscreator news
Run the deterministic safety gatescript
What it blocks
DeathCrimeDrugsMedical emergencyMocking disability
A script holds this line, not a model
Held, and reported. Never dropped silently
Judge authored content versus spectacle23 → 12
A joke, or just a spectacle
23came in
11cut here
12still in
Did someone write it, or did someone happen to film it
eleven sent back on category alone
Test whether it is worth sharing12 → 4
Would I forward this
12came in
8cut here
4still in
Cute, warm, odd and rare all fail here.
Remove content built on embarrassment4 → 3
Is it unkind
4came in
1cut here
3still in
Someone's body, their embarrassment: out, however funny.
Finish the edit and platform copyfinished asset
Final cut
Platform copy APlatform copy BSource credit retained
A person watches and decides3 → 1
Finished asset
A person must play and watch itFinal reviewTitle and footage checked togetherAttribution chain completePerson makes publish decision
3handed over
1published
0autonomous
Yesterday's recorded rejection reasons become today's starting context.
Titles, review states, and the feedback-loading rule come from the live system; the counts are from its first real review pass.
System design
Critical controls
- 01
Feedback becomes system memory
YesterdayRejectionsPersistSystem memoryTodayLoad firstEach run starts with fresh model context, so the learning cannot live in chat history. I persist the reasons real work was rejected and load them before the next run.
- 02
Safety is deterministic where it can be
ScriptFixed safety rulesModelContext and contentPersonWatch, then decideHard exclusions are enforced by code the creative model cannot loosen. The model is used where context and taste matter, not where a stable rule is more reliable.
- 03
Publishing remains a human decision
TitleRule passedFootageNot yet seenPublish gatePerson watchedA title-only filter cannot see every risk in footage. Human review is therefore not ceremonial approval; it is the only step that has actually watched the final video.
In operation
Operating results
- 5/day5/dayvideo and image-and-text deliverables
- D-1→DD-1→Dhuman feedback becomes system memoryYesterday's review improves today's selection
Titles, review states, and the feedback-loading rule come from the live system; the counts are from its first real review pass.