How to Make an AI Short Film: From Shotlist to Export

Make an AI short film by turning a brief into a shotlist, approving character and location references, comparing takes and cutting the selected sections on a timeline. ChatCut’s historical C3 run exported a 14.625-second scene with 6/7 MUST checks passing. Framing still failed and preferred audio remained uncertain. It demonstrates a small scene workflow, not a validated whole-film or one-click production guarantee.

Verified facts and their scope

Last verified: October 28, 2026. Measurements cite the C3 Santiago report named in Sources, from a local integration run on October 27, 2026. Deployed production was not retested for this page.

Choose a model with the capability matrix and prepare inputs with the reference guide.

How does a script become a shotlist?

Start with a scene brief: what changes, who is present, the visual tone and the dialogue. Turn each story beat into a shot with a subject, action, camera family, eyeline and duration target. Keep a scene-level Look record for the lighting and palette so a new prompt does not redefine the scene.

The measured C3 example used three shots in order: a two-person office wide, a psychologist single with the line, and a silent Santiago single. The target shot lengths were 104, 88 and 163 frames at 24 fps. This gave review something concrete to judge beyond whether the footage looked attractive.

Write acceptance criteria before generating. A two-person wide should show both people in distinct occupied chairs. A close-up should identify who is speaking and where they look. Separate framing requirements from identity and color requirements.

How do casting locks carry across shots?

Approve one appearance reference for each character and a separate location plate. C3 generated its own candidates and selected one-face character sheets. The same locks were reused in every video; source tutorial frames were only an external review target.

Use stable names in the shotlist and bind each name to the approved reference. On a character’s return, check against the original lock. Changing the reference to an already drifted take can compound the error. Manual inspection remains necessary even when the same references are reused.

How are takes reviewed and harvested?

Review each shot within each take. One take can have the best wide while another has the best single. In C3, four identical-prompt takes were compared before the selected sections were assembled. Two later targeted wide repairs shared the same camera defect and were rejected.

How is pacing measured in the export?

The C3 edit used the wide from take 3 and both singles from take 2. The final shot lengths were 104, 100 and 147 frames: 4.333333, 4.166667 and 6.125 seconds. Their differences from the source targets were 0, +12 and −16 frames, within the run’s ±24-frame allowance.

The final export contained 351 frames at 24 fps, so its duration was exactly 14.625 seconds. Cuts fell at frames 104 and 204. Adjacent-frame inspection confirmed two clean joins without an old-shot flash. Generated raw takes had different timings; the timeline edit produced the delivered rhythm.

Which production checks remain manual?

C3 passed bounded character matching, geography, dialogue ownership, warm appearance and pacing. The reviewer had prior project context, so passing identity controls does not establish an independent biometric or all-character guarantee.

The camera requirement failed. The wide stayed lower/frontal rather than high/oblique; the psychologist shot’s cant ran in the opposite direction. Rotating the wide changed tilt but could not change the original viewpoint. Preserve this failed check even if the overall scene feels usable.

Listen to the exported audio separately. C3 recovered the intended line, but room tone and the absence of music were not independently verified. A transcript, waveform or lack of a music asset cannot establish the complete mix. Review sound, permissions and delivery requirements before sharing your own film.

Budget by approved batches and compare estimates with settled usage. C3 required casting, location candidates, full takes and rejected repairs; one successful export does not imply that every paid take was usable. The credit policy owns current rates, while workflow failures explains how to handle rejected results.

Sources

• ChatCut C3 Santiago short-film acceptance report, October 27, 2026, reviewed October 28, 2026: historical local integration run with original casting, four full takes, selected-section editing and measured export. The internal report and media are not publicly hosted. It validates neither a whole multi-scene film nor current production availability.
• Santiago tutorial scene, reviewed October 27, 2026: attributed external framing and pacing target; never a generation input. Tutorial content is not evidence for ChatCut’s measurements.
• Seedance generation API, accessed October 27, 2026: native model input modes and timing constraints.