The product brief as a prompt.

The first time I watched Rumana apa run a 4am crew call, the unit was rigged on a Bakshi Bazar lane and the December air was already thick with the diesel of three minibuses idling outside Star Kabab. She was 42 then, a script supervisor on her ninth feature, and her clipboard carried two pieces of paper. The screenplay sat on top, forty pages thick, written by the writer two months ago in a flat in Lalmatia. Underneath was the day's call sheet, a single page on carbon copy yellow stock with five scene numbers down the left margin and a weather strip across the top. Iqbal bhai, the DOP, walked across the lane, took the call sheet from her clipboard, and read it under the cracked sodium lamp by the corner pharmacy. He never opened the screenplay.

I have been thinking about Rumana apa's clipboard for the last year, because the product briefs I see being written in agentic teams have started to behave the same way. There is still a long document somewhere, a PRD that took two months and three rounds of review, parked in a folder nobody opens after the kickoff. The thing the engineer is actually building against is a much shorter artifact, written for a model rather than a person, that gets revised between takes. The collapse the field is living through, in 2026, is that the brief is now what generates the implementation. The brief is therefore the spec, the eval, and the documentation folded into one page on yellow paper.

The shape of the argument has been visible for a while. Marty Cagan was calling for the death of the PRD back when empowered product team was still the operative phrase, on the grounds that the document had stopped being useful the moment teams started discovering customer behaviour iteratively. Andrej Karpathy made the technical version in his Software 2.0 essay; the program is no longer the artifact, the dataset is. Geoffrey Litt has spent two years arguing that prompts are artifacts for programming by end users, worthy of the same craft and version control we used to give to source files. When the brief is what produces the build, you stop writing it for the steering committee and start writing it for the model.

What that brief looks like, when somebody means it, is shorter and stranger than the PRD it replaces.

The five things on Rumana apa's call sheet are, roughly, the five things on a brief written for an LLM, in the same order.

The first move is to name the constraint. Rumana apa's call sheet for Day 14 has a single line at the top that reads wide lens stays wide, in indigo pencil, underlined twice. The DOP knows the screenplay calls for the camera to dolly in on the rickshaw at the end of Scene 14A. The constraint says he is not switching to telephoto for a closer grab, even if the boy on the rickshaw turns out to have a face you want to be closer to. A brief written for a model carries the same line at the top. The agent answers in JSON and never in prose, or the recommendation cites at least one source from the indexed corpus. The constraint goes first because everything else on the page is downstream of it, and because the model, like the DOP, will improvise around it if it is not visible from across the lane.

The second move is to declare the failure mode. Rumana apa's sheet for Scene 14B reads if the boy looks at camera, do the take again, which is not a hope but a rule that changes what gets shot. The corresponding move in a brief for an LLM is the plain English line that the model may not invent a refund amount the order system has not returned. Or that it must escalate when the customer is shouting, instead of trying to console. The failure mode named in advance is the one the eval suite catches before it ever reaches a customer.

Sc. 14A · take 3

35mm · f/2.8

A brief that does not name what failure looks like is a brief the model is going to fail in a hundred small ways nobody noticed in advance. It is the same as a take where the boy looks at camera. You only ever realise it was unusable much later, in the edit suite at the Mohakhali office at 11pm on a Tuesday.

cam asst initialled

✓

the line

The third move is to embed the eval. The call sheet for Scene 21, the porter at the ghat, reads three takes minimum; one has to land before the porter coughs. The cough is the test. Rumana apa wrote it in the margin because there is no point shooting forty takes if you cannot tell which one you are going to print. A brief written for a model carries the same. The eval rows live inside the brief itself, not in a separate test plan stapled to the back. When the eval lives in the brief, every revision of the brief revises what working means, and the team stops arguing about whether the new version is better than the old one because the eval row in row 14 of the suite decides.

The fourth move is to say what NOT to ship. Rumana apa's sheet for Scene 22 reads do not roll on the bystander who keeps wandering into frame, with the bystander's name in parentheses because the PA has been failing to keep him out for three days. A brief written for a model carries the equivalent. Do not promise a discount the order system has not authorised. Do not narrate the model's own reasoning back in chain of thought form. The negative space on the brief is the part the steering committee version leaves out because it feels uncharitable to write down what the team is choosing not to do. The team that does not write it down ships the bystander in the frame.

The fifth move is to cite three precedents by hand. Rumana apa's sheet for Scene 25 reads match the framing of the 1987 Tareque Masud kid and river shot. It is one line, and it carries more direction than three paragraphs of compositional notes would have. The DOP knows the shot, knows the lens, knows what was not in the frame, which is half of the direction. A brief for a model uses the same trick. Three or four exemplars of a great answer, picked by hand from real traces, beat a description running a thousand words of what good means. Robert McKee writes in Story that the working screenwriter's shelf is half craft books and half scripts of films he loved; the brief written for a model is the same shelf, scripts and craft notes side by side.

The collapse those five moves describe is that the brief is no longer a translation step between the steering committee and the engineer. The brief is the artifact the model reads. The PRD still gets written because somebody has to think the long thoughts. But the call sheet on Rumana apa's clipboard is what gets shot. By the time the camera assistant initials the bottom of the sheet at 6:42pm, the day is in the binder and the next morning's sheet is already underneath.

The thing that makes this work, in both registers, is that the call sheet is allowed to be wrong on Tuesday and corrected by Wednesday. The brief is allowed to be wrong on the first prompt and rewritten by the third. The screenplay never moves; the call sheet is the part of the document that gets to learn. When I sat with Rumana apa in the unit van on the way back across the Buriganga that morning, the wind smelling of fish and outboard fuel, she had already pencilled the next day's plan, having lifted Scene 17 from the day she had just shot because the boy who plays the grandson had come down with fever, and rather than throw out the screenplay or pick up the phone and ring the writer back in Lalmatia, she simply wrote tomorrow's call in its place.