Before you start
A spreadsheet or coding agent, three fictional shots and approved reference art.
Why this lesson exists
This lab adapts the “Treat AI Video Like Software” project write-up into a practice build. The time is an estimated first session, not a promise to finish a production system. Use the public repository as a reference when available; the exercise can be built with original sample content.
Do the exercise
Define the practice version
Track shot ID, purpose, prompt version, reference files, output file and review status. Start with manual generation rather than provider integrations.
Build step 1
Write character, location, and style canon.
Build step 2
Define each shot by purpose before prose prompting.
Build step 3
Generate a first frame and last frame.
Build step 4
Review and approve references.
Build step 5
Carry approved frames into the next shot.
Build step 6
Build the storyboard before paying for final video renders.
Run the experiment
Revise shot two while preserving its earlier prompt and output. Confirm shot three points to the approved reference, not a rejected version.
A prompt to adapt
Replace the bracketed parts with your own practice details.
Work in a disposable practice project. Explain any setup requirements before changing files. Build one small step at a time and show how I can check it. Convert this music-video concept into a canon document with immutable character traits, wardrobe, locations, palette, lens language, lighting rules, prohibited drift, and continuity checks. Compile a standalone image-generation prompt for shot 07. Use the supplied canon and shot intent. Include subject, action, environment, composition, camera, light, color, continuity anchors, negative constraints, and a short quality checklist. My first-version boundary: Track shot ID, purpose, prompt version, reference files, output file and review status. Start with manual generation rather than provider integrations.
Run this experiment
Revise shot two while preserving its earlier prompt and output. Confirm shot three points to the approved reference, not a rejected version.
Check your result
Use evidence from your output. A confident explanation from the AI is not enough.
- Every output can be traced to a prompt and references.
- Rejected versions cannot be mistaken for approved ones.
- A new collaborator can find the next action.
If it isn’t working
If automation retries a paid job after a timeout, inspect its existing job ID before submitting again. Keep creative review separate from provider success status.
Optional. Progress stays in this browser.
Where this came from
Public project repository ↗. The practice lesson is an adaptation, not a verbatim transcript. About the sources.
Prepared September 2026. Tools and interfaces change; use current official setup instructions. Session lengths are estimates.