Case study 03 · Creative direction and AI workflow

I taught an AI videosystem my taste.

Afterlight Studio is a video-production system I designed. It makes short film-edit drafts. I rate every draft from 0 to 10 and pin notes to exact frames. My notes become rules, and the next drafts have to follow them.

My role
Creator and creative directorPersonal project
Timeline
Oct 2026 to nowData: Oct 26 to Nov 16
Budget
$0 extraFree and open tools only
Skills
Creative direction, content testing, data-driven iteration, AI workflow design
3.1 → 7.7average draft rating, out of 10, in three weeks
138drafts rated across those three weeks
1 → 10my rating, Oct 29 draft vs. Nov 18 draft
286open AI models that passed my tests

Not postedNothing has been posted yet. No views, followers or revenue are claimed.

Before · Oct 29, 2026

1 /10

This is bad. Editing makes no sense, no storytelling, and too short.

After · Nov 18, 2026

10 /10

Wow, this is amazing. Just take the caption out and this is ready to post.

My own notes, three weeks apart. The drafts themselves aren’t shown because they use film clips I don’t own.

The problem

My notes on one draftnever reached the next.

Editing short-form video well is slow. AI tools can make drafts fast, but they don’t remember what you told them last time.

The early drafts averaged 3.1 out of 10. With no memory, each new draft could repeat the last one’s mistakes, and my notes went to waste.

What I needed was a system where every note I write changes every draft after it, so quality keeps building.

Where a note goes

BeforeNote → next draft ignores it → same mistake
With the loopNote → rule → checked on every later draft

My role

I judge the drafts.The system makes them.

The AI does the cutting. Deciding what good looks like, and proving the system is getting closer to it, is my job.

What I did

  • Designed the loop: draft, rate, note, rule, next draft.
  • Rated every draft from 0 to 10 myself, and wrote why.
  • Pinned notes to exact frames, so each rule points at the moment it fixes.
  • Decided which notes become rules and which were one-offs.
  • Set the limits: $0 extra budget, and a 7-part scorecard where a fix only counts with a before-and-after number.

What AI tools did

  • Made each draft: picked shots, music and timing.
  • Turned my rules into checks that every new draft must pass.
  • Ran 1,145 model tests; 286 open models passed. A router picks one for each job.
  • Wrote the code. I direct the work; I don’t hand-write it.

03 · Process

Every rating feedsthe next draft.

The loop is simple on purpose. The hard part is writing notes clear enough to become rules.

Step 2 is the only one I do by hand, and it’s the one that matters.

Rate everything, including the bad ones.

A 1/10 with a clear reason teaches the system more than a 7 with no note.

Notes point at exact frames.

A general note like “too short” is hard to act on. A note pinned to the frame where the story drops becomes a rule the system can check.

Trade-off
Slower reviews. Each draft takes real attention.

Fewer drafts, higher bar.

The week of Nov 9 had 85 drafts. The week of Nov 16 had 11. As the rules took hold, volume dropped and the bar went up.

$0 extra.

Free and open tools only, and a fix only counts when it comes with a before-and-after number.

Step 2, on one real draft

Nov 10 draft: 37.5 seconds, 16 cuts, rated 4/10.

I pinned 9 notes to exact moments. The film clips aren’t mine, so the 12 frames below show only their light: how bright each part of the picture was, measured from the draft. The cuts, times and notes are real (spelling fixed).

Keep Cut

Drag the film or tap a pin · arrow keys step · Page Up / Down jump between notes

  1. Frame 01 · 0:00.5 · Shot 1 of 16

    CutNote 1 of 9 · 0:00.1–0:01.9

    Weak intro. It doesn’t make much sense.
  2. Frame 02 · 0:03.8 · Shot 2 of 16

    KeepNote 2 of 9 · 0:02.6–0:04.0

    I like how the lyrics match here: they’re walking together through a foreign town, so it makes sense.
  3. Frame 03 · 0:07.1 · Shot 4 of 16

    CutNote 3 of 9 · 0:06.2–0:07.5

    Out of place. Remove it; the editing seemed bad for this.
  4. Frame 04 · 0:10.5 · Shot 7 of 16

    No note on this frame.

  5. Frame 05 · 0:13.8 · Shot 8 of 16

    No note on this frame.

  6. Frame 06 · 0:17.1 · Shot 9 of 16

    CutNote 4 of 9 · 0:16.1–0:17.5

    Remove this. It’s not part of the show.
  7. Frame 07 · 0:20.4 · Shot 11 of 16

    CutNote 5 of 9 · 0:18.0–0:19.6

    It doesn’t make much sense to just show a road.
  8. Frame 08 · 0:23.7 · Shot 12 of 16

    CutNote 6 of 9 · 0:22.9–0:26.9

    Think more scenes of longing for her instead of this.
  9. Frame 09 · 0:27.1 · Shot 12 of 16

    No note on this frame.

  10. Frame 10 · 0:30.4 · Shot 13 of 16

    CutNote 7 of 9 · 0:27.6–0:31.7

    Same here: more scenes of longing for her.
  11. Frame 11 · 0:33.7 · Shot 14 of 16

    KeepNote 8 of 9 · 0:32.0–0:33.4

    This clip felt cut fast. It’s not bad footage; I wish it stayed about 4 seconds longer.
  12. Frame 12 · 0:37.0 · Shot 16 of 16

    CutNote 9 of 9 · 0:36.2–0:37.5

    The ending makes no sense.

Overall note: We need to fix this. Notes like these are what the rules below are written from.

The work

How one notebecame a rule.

The rules live in plain words, so I can read them, argue with them and change them. These four got the Nov 18 draft to a 10.

Research the story first.
Pick shots that tell it.
Let the edit linger.
End on a fade to black.

The Oct 29 note said “no storytelling” and “too short”. The rules behind the Nov 18 draft answer both: research the story first, and let the edit linger.

That’s the skill this project trains: turning a gut reaction into an instruction a person or a machine can follow.

05 · Results

My average rating morethan doubled in three weeks.

Average rating of film-edit drafts, by week. Every rating is mine.

Oct 26 · 42 draftsNov 9 · 85 draftsNov 16 · 11 drafts
+4.6points on a 10-point scale, from the week of Oct 26 to the week of Nov 16
138drafts rated; bars are averages, each draft got one rating from me
1,145model tests run; 286 open models passed

Not claimedViews, followers or revenue. Nothing has been posted. The ratings measure my judgment, not an audience’s yet.

What’s next

Next: test my tasteon a real audience.

  • Post with footage I have the rights to, starting with Brivan’s own content.
  • Track watch time and follows per post, next to my rating for the same post.
  • Compare the two. If my 9s don’t beat my 6s with viewers, my rules are wrong and get rewritten.

What I learned

Good taste can bewritten down.

A note that becomes a rule keeps working on every draft after it. The same habit works on a team: specific feedback, written once, saves everyone from hearing it twice.