A director that rereads every film before it is shot

Inside the website, the planner writes a storyboard in one pass, from the page's text, its ranked subjects and a short suggested brief. Around it sits a stack of careful rules, each one guarding against one defect: a reviewer that rewrites a faulty scene, a minimum share of real recordings, at least one real action, the walk, the designed cards.

Every rule sees one piece; none of them rereads the whole film with judgment. What slipped through was exactly what a person repairs in a Claude session: a click on a button that does nothing, a scene about a page the film never shows, a card that repeats the voice, a claim the page does not make.

In Claude, that repair is a simple loop: read the storyboard with get_storyboard, fix it with update_storyboard, read it again. The question was whether the website could follow the same loop.

It can. The director runs that loop inside the planning step. Once the plan is written and reviewed, a stronger model, Gemini 2.5 Pro, is handed exactly what a Claude session is handed:

• the storyboard, as an assistant would read it;
• the subjects the probe ranked, with what the walk saw each button do: open a page, open a panel, nothing, or a sign-in door;
• the site's pages and the brief;
• the same instructions our MCP gives assistants: the storyboard checks, the repair guide, the craft rules.

It answers with the same edit operations an assistant sends. The same validator checks them, the same service applies them, and the recorder films the result. There are two passes at most: the second rereads the repaired storyboard and answers with nothing when it is right.

A model that edits someone's film needs limits that do not depend on its good behaviour. So the guard is written in code, and it drops any operation that would:

• spend what the person did not ask for, such as a presenter or AI visuals beyond the film's AI setting, so a free film stays free of AI;
• change the person's voice, music, look or pronunciations;
• film a button the probe did not rank, or a page the site does not have;
• touch a scene the person wrote themselves, or reorder scenes without naming each one exactly once.

And the director never fails a film. If the model errs, refuses or runs out of time, the storyboard simply stays as the review left it.

We ran the director on test films before turning it on, and two lessons came straight from that bench.

The first was about refusals. Our MCP refuses a batch of edits whole, which is right for an assistant that reads the refusal and tries again. Inside the pipeline nobody reads it, so a single slip could set aside a whole pass of good edits. The director now checks each edit on its own, keeps the valid ones, repairs the common slips, and hands a refusal back to the model once, as a careful assistant would.

The second was about length. The director's repairs were first treated like a person's own words, which our length fitter never touches, and the bench films came out short. Its edits are now marked as the machine's, and the length fitter runs after them.

We benched five sites with no brief, the way a visitor arrives:

• on pagetovid.com, 8 repairs: the soft opening became typing an address into the field, a card became the recording of the real section, and the film ended on a real click;
• on nyna.ai, 5: a click into a sign-in page became a spotlight, a question typed into the chat was followed by its answer, and claims the page does not make were rewritten from what it shows;
• on aifr.ai, 5 in the first pass and 3 in the second;
• on a construction company's vertical film, 4, though it still opens on a text card;
• on a bike shop's tutorial, a messaging-app click and a click that did nothing became spotlights.

The director adds one to three and a half minutes before filming; recording, voice and render are unchanged. As of today it is on for every film, with one switch to turn it off. To see the gap it is closing, compare the three films on /difference.