Compare approved script versions

An agent here means an AI set up to do recurring work for you, such as reading new script pages and reporting what changed. A proactive agent starts that work on its own when something happens, without waiting for you to type a prompt.

Before an agent does anything, it needs a remit: the job it is allowed to do. It also needs a hard line: something it must never do, however helpful that might seem.

Ted Sarandos, Netflix's co-chief executive, described this pairing on Prof G Markets in June 2026. Writer friends of his use Anthropic's Claude as "a writing partner". They give it everything they've written and test ideas on it all day, but they forbid it to write for them. That is the remit (be a sounding board) and the hard line (never write the pages).

Every agent in this course follows that model. For a script agent, the remit might be: read approved versions and explain the differences. The hard line might be: never change the approved baseline. The approved baseline is the version of the script that the production is officially planning from.

Richard Chuang co-founded Pacific Data Images and built the pipeline behind Shrek. On Machine Cinema in March 2026, he described a request from a large studio. It wanted an automatic pipeline that would update the script, the scene breakdown and the generated content on its own, then release, take feedback and revise itself. He called his own sketch of it rough and warned against taking such a pipeline for granted.

That warning shapes this course. No recording shows production agents doing this work alone, so you will treat every agent here as a supervised exercise. The agent prepares, a person reviews, and you test whether it stops when it should.

In practice, your revision agent will not "keep the script up to date". It will read two approved versions and write a report for someone to check. That is a smaller job, but you can verify it.

An agent brief is the written instruction that defines an agent's job. Start by naming the role, for example Sequence Desk. A role name tells everyone what the agent is for. It also makes it obvious when the agent strays outside that job.

The brief then sets four edges:

• What wakes it. This can be an event, such as a new file arriving in the approved script folder, or a schedule, such as every evening at 7 pm.
• What it may change. For a revision agent, this is only its own comparison reports in a draft output folder.
• What it must never touch. That means the approved baseline, the picture, schedule and cost decisions, and performer likeness.
• Whom it asks. The script supervisor answers questions about which version is current.

The brief also lists its connections (the apps and folders it can reach), its report format, its review points (where a person checks its work) and how it is paused or retired. Retiring an agent means switching it off for good. You'll write each of those in the steps that follow.

You can build this agent in either of two tools, and the course treats them as equal options.

• Dots comes from OpenAI. OpenAI describes it as a proactive agent that runs on its own cloud computer, is connected to your apps, and checks with you before sending, paying, signing or deleting.
• Claude Duet is Anthropic's equivalent.

Choose on practical grounds rather than on features. Which tool does your production already have access to? Which one has your producer or IT lead approved for script material? If the answer is "Dots on the production account", use Dots.

Both products are new. Read the tool's current documentation for how its triggers (what starts a run), schedules and permissions work, rather than assuming. If the documentation doesn't say whether the agent can watch a folder for new files, use a schedule instead and note that in your brief.

Write your choice and your reason into the brief. For example: Claude Duet, because the production already approved it for development documents.

Next, decide how long the agent works.

A single run handles one revision pass. Example: the blue pages for scene 42 arrive, the agent compares them with the previous approved version, you review the report, and the job closes. Nothing keeps running afterwards.

A continuing brief keeps following new pages over weeks, for example through prep. That is more useful, but it also has more ways to go wrong without anyone noticing. Before it starts, it needs three things:

1. A named human owner. This is a real person, such as the script supervisor or the coordinator, who answers for the agent.
2. A review cadence. For example, every report is checked the next morning before anyone acts on it.
3. A stated end condition. For example, the brief ends when picture is locked or when the owner retires it, whichever comes first.

If you can't fill in all three, start with a single run. You can turn it into a continuing brief once a few reviewed runs show that it behaves as intended.

A connection is a link that lets the agent reach an app or folder. Connect only what the exercise needs.

For Sequence Desk, that means two folders:

• Read access to the approved script folder. The agent can open the pages but cannot edit them.
• Write access to a separate draft output folder, where its reports go.

Keeping these apart is your first protection for the baseline. Because the agent can only write to the draft folder, it can't overwrite the approved script even if something goes wrong.

Leave everything else out unless its owner approves the connection. That includes:

• unreleased story material beyond the pages being compared
• performer information such as contact details or likeness references
• private crew records

An agent that is only meant to compare pages has no reason to see the cast list or the crew's phone numbers.

Write the connections into the brief exactly. For example: Read: /Approved Script. Write: /Sequence Desk Drafts. Nothing else.

Now put the brief to work. Your revision agent compares two approved script versions and explains the difference scene by scene. It runs when a new file arrives in the approved folder, or on its schedule.

In self-paced practice, the tutor supplies the pages and plays the script supervisor. Use the course's scene forty-two, or a small non-confidential sequence of your own.

The example: the previous approved pages, then the blue revision. In the blue pages, Nora now sets the house keys on the porch rail before she tells her brother she's selling the house.

A good run follows a fixed order. The agent wakes, confirms which two approved versions it is comparing, compares them scene by scene, separates what the script says from what it infers, writes the report to the draft folder, and waits for review. It edits nothing in the approved folder.

Then check the report against the pages yourself. Did it find the keys? Did it invent a change that isn't there? Missing a change is a failure. So is inventing one.

An agreed report has a fixed format that you and the reviewer settle in advance, so every run can be checked the same way. Sequence Desk's report has four parts:

1. Versions compared. The title, date and page color of each version. Example: previous approved pages vs blue revision, with dates as printed on each.
2. Changes by scene, in plain language. Example: Scene 42: Nora now sets the house keys on the porch rail.
3. Stated versus inferred. What the script actually says, kept apart from what the agent infers. Stated: the keys are set on the rail. Inferred: the keys may need to appear in a later scene. That inference is only a question for a person. It does not create a new requirement.
4. Needs review. Anything uncertain, plus who should look at it.

Keeping the third part separate stops plausible guesses from turning into production requirements.

Next lesson, you'll protect the baseline when the pages misbehave, for example with duplicates and drafts that have the wrong date.