Markdown tracking ties each AI video run to its inputs

The guide builds its method around a simple failure mode: saving only the finished video loses the link between an input and its result. To prevent that, it separates the folder into assets for source material, runs for per-generation Markdown records, and outputs for the downloaded videos, and it assigns a new ID even when regenerating under the same prompt so that a changed-condition run and a same-condition rerun stay distinguishable. Versioning is handled the same way — a trimmed or brightened asset is saved as v2 rather than overwriting the original, so older records still point to the input that actually existed at the time.

The record itself splits into observation and judgment. The guide says to note what was seen — outline, printed text, motion, and the moment it was checked — then whether the run is adopted, regenerated, or held, and which checklist items were met, because observation alone does not identify whether the prompt, the image, or the generation process caused a change. It also warns that a stated length like 5 seconds is a pre-generation request; the actual output must be measured separately. Likewise, if no model name is shown on screen, the record says "not displayed" rather than filling in a guess.

At the end of a project, the guide suggests appending the adopted generation IDs and reasons to project.md and retaining records for unused outputs so a similar fix later has a starting point. It is explicit about the limits: having the full input history does not guarantee the same video can be regenerated. What the record can trace is the material used, the instructions sent, the output obtained, and the adoption decision — with review and logging still done by a person, including when AI generates the video automatically.