The Method
The authoring method has three stages, run in order, once per language:
- Structure the material into a narrative arc. The classical four-part shape — setup, development, a turn, resolution — applied to whatever actually happened.
- Unpack it so a reader or listener can follow. Explain, expand, give the abstractions something concrete to stand on.
- Make it taste good, fit the rhythm, refine again — as a loop, not a pass.
Stated like that it sounds unremarkable. What makes it worth a lesson is the failure mode: in practice stage 2 is the one that silently disappears, and stage 3 is very good at concealing that it did.
A Beat Table Is Not an Arc
The first trap is mistaking a plan for a structure. The pipeline has a table saying what the screen carries per beat, and it is easy to treat that as stage 1 having happened. It hasn't. A beat table is a screen plan: it says what is visible, not what is being argued or in what order the understanding builds. Stage 1 still owes an arc — and it may discover the material has more than one turn, which is exactly the shape worth slowing down for.
The second trap is that stage 2's output has no obvious artifact, so it is easy to believe it happened. The rule that fixes this is blunt: stage 2's artifact is prose, and it must exist as a file. Flowing prose, told to somebody. Not an outline, not segments, not annotated bullets. If no such document exists, stage 2 did not happen — and that is checkable, which is the whole point of demanding a file.
Downstream Constraints Are Banned From the First Two Stages
Subtitle widths, segmentation rules, language-uniformity requirements — all of these belong to stage 3. Admitting them earlier is how an invented constraint starts driving the prose.
The pilot demonstrated this precisely. It compressed itself to a self-imposed subtitle budget of 34 characters when the real limit was around 41 — and its widest line ended up using 56% of the width available. It had squeezed its sentences to fit a number that did not exist, and the squeezing is what produced the compressed, unexplained prose that got rejected. That is worth remembering as a general shape: an imagined constraint applied early does more damage than a real constraint applied late, because nobody ever checks whether it was real.