INSIDE THIS AI SKILL
The craft behind its decisions.
These are the principles and procedures the skill applies. You can understand the reasoning without having to perform every production task yourself.
Retention is a structural property, not a post-production fix.
An episode that loses people at forty seconds usually has a plan problem at forty seconds. The edit can sharpen it. It cannot invent a reason to keep watching.
01 / Retention editing
The opening three seconds
The first frames arrive before any context. No title, no recap, no patience. What has to be true is that a stranger scrolling sees something specific enough to want the next second.
The reliable failures are all forms of delay: a logo, an establishing shot, a line of setup dialogue, a slow fade. Each of them spends the only currency the episode has before it has earned any.
- A logo or a title card
- An establishing shot of a place
- Setup dialogue explaining who people are
- A fade from black
- A recap of the previous episode
- A face, mid-reaction
- A line that only makes sense if you keep watching
- An object the audience already knows the meaning of
- A physical action already in progress
- The consequence of last episode's ending, immediately
02 / Retention editing
Where completion is actually lost
- 01Find the drop in the curve, not the average
A completion percentage tells you how many left. The curve tells you where, which is the only actionable version of the information.
- 02Look at the shot immediately before it
Drops usually follow a moment where the episode stopped asking a question - a scene resolved, a tension released, a beat held too long.
- 03Ask whether the scene has a job
The most common cause is a scene that exists to convey information the audience could have inferred.
- 04Cut, do not shorten
Trimming a scene that should not exist produces a shorter version of the same problem.
- 05Check the ending placement
If people are leaving in the last fifteen seconds, the episode probably resolved before it stopped.
- 06Compare across language tracks
The same episode, same picture, different track. A divergence there is about the track, not the story, and it is a cleaner comparison than most tests ever get.
03 / Retention editing
Reading the numbers honestly
Questions about this skill
Can I shape the edit without knowing editing techniques?
Yes. Describe what the audience should understand or feel. The Retention Editing Skill helps your AI Director propose an opening, edit rhythm, reaction holds and an ending transition. It makes those craft decisions explicit, while audience response still has to be measured after release.
How long should an episode be?
60 to 120 seconds for our serialized launch format, with a 150 second hard cap during validation. Within that, shorter usually wins when the alternative is padding, and the cap exists so cost per accepted minute stays forecastable.
Should I recap the previous episode?
Almost never at the start. A recap spends the three seconds that decide whether anyone watches, on information the returning viewer already has. If context is genuinely needed, deliver it inside the action.
Does the cover frame matter?
On a vertical feed it is part of the hook, which is why cover frames are a selected element of the delivery package rather than a grab from the first frame.
Can I fix a retention problem in the edit?
Sometimes, at the margins. A drop caused by a scene with no job is a plan problem, and the cheapest place to have fixed it was the animatic. That is the argument for boarding and timing the episode before rendering it.
How many episodes before the numbers mean anything?
The pilot arc is five episodes, which is enough to see a shape across a release cadence rather than a single upload's luck. Continuing, revising or stopping are all legitimate conclusions from it.