AI Agents and Filmmaking

There is a lot of “AI slop” out there, and its going to increase. Isn’t that bad? So why am I talking about AI agents and Filmmaking?

An AI influencer I follow (Nate Jones) once said (paraphrased) that you can think of AI like a second brain. Its hard for humans to keep too much information in your head at once. That is where AI can help. Not to replace the human, but to extend their capacity. That is my interest. Using AI to help humans create, not replace them. AI slop comes from removing the human in the loop.

For example, my goal is to create several YouTube series. If successful, I will be running them in parallel over an extended period of time. (Years! Well, in my dreams at least! Lol!) How then to make sure inconsistencies don’t creep in?

This is an example of where AI can help. I will write the script for new episodes, but I then plan to run AI over the script to do a consistency check against all previous episodes (or summaries of those episodes) to make sure I don’t say contradictory things. I can also feed in all previous dialog for the characters in the episode - are they using similar language and tone, faithful to the original vision I had when first creating the character? They are allowed to evolve over time, but it should be intentful, not accidental.

So when I started planning to add agent support to my backend, the first thing I thought through was how should humans provide the information the AI needs. I want to understand the writer’s intent, not just the final content of scripts.

To help with this, I have been building a timeline view with multiple swim lanes.

Time (in episodes) goes across the page, you can define swim lanes down the page to represent different concepts. You can drag out cards both across episodes and within episodes (if you zoom in).

In the screenshot, the first swim lane I used for seasons of a series. The second is story arcs that span episodes (if they exist). Then finer grain you get into story beats within episodes, which may carry over between episodes, especially if you end episodes with a cliffhanger. I am also considering character development swim lanes, to keep track of character development beats (as distinct from story beats). A character development beat might be “character A realizes they are in love with B”. Swim lanes can also be used for sub-plots, to note real world time, and so forth. Everything can be labeled as you wish.

Of course shortly after completing a first cut at implementation, YouTube was watching and suggested the following video on Plot Grids to help with writing a book.

Its the same concept, just organized slightly differently.

• A story grid is basically a spreadsheet.

• Time goes down the page, where time is measured in chapters.

• Columns in the example early in the video (from a Harry Potter book) were:

  • Chapter number

  • Month it happened (its a school year, so time matters)

  • Chapter title

  • Plot

  • Extra columns per subplot.

Its a useful way to organize and group concepts over time.

How does a timeline evolve? My plan is to iteratively refine the content

• Start with high level concept of seasons. I may never get to season two, but it helps me capture ideas for the future but leave them out for now.

• Rough out story arcs. I plan to to short episodes because of limited time. I would rather publish episodes a few minutes long each week rather than long episodes but take a quarter to finish on episode.

• In parallel, start thinking through character beats. What are significant personal events and what ordering dependency is there between them?

• Start roughing out story beats.

• Then flesh out a few episodes in advance (not all of them).

• Episodes are broken down into a few scenes.

• Then take all the concepts from different swim lanes and provide them to the writer (me!) so they remember everything that should happen in an episode. Some might get pushed to later episodes if they don’t fit. At all times, the overall structure can change. Nothing is finalized until an episode is released.

• Once released, all information about previous episodes is kept online for reference.

What’s next?

• Flesh out the timeline functionality (mostly done).

• Provide contextual information from the swim lanes to AI agents, so the author can build up agents to review new episodes against history for consistency.

• I also want to implement an “undo” mechanism for changes (currently there is no undo feature).

• Then I want to let AI agents make changes to the content. Implementing “undo” first is a safety mechanism against over-enthusiastic agents.

None of the above is about screenplay to AI video. It is all about tools to help humans write engaging stories and get feedback from AI to improve them. Once there is a good engaging story, then I can get back to my original focus on screenplay to AI generated video.