A producer wants a synopsis by Friday, then a treatment. A screenwriter goes from idea to document in five days, with an AI that asks questions instead of writing, and without her film getting around.
Monday, 9:10 am. Léa hangs up with her producer. The call lasted four minutes and comes down to two sentences: the idea has gone down well, she needs a three-page synopsis by Friday, then a first treatment of around twenty pages for the development funding application. On the table is the notebook where the idea was born: half a page of notes about a mountain hut warden who, during a three-day storm, takes in a hiker she thinks she recognises. The rest is arrows, two lines of dialogue and a circled question mark.
Léa is a fictional screenwriter, but her week looks like that of many writers. Here is how she gets through it with AI, one request after another, without losing control of her film.
The synopsis tells the whole film, ending included, in a few pages, in the present tense and without dialogue. It says who the character is, what they want, what stands in their way and how it ends. The treatment goes further: it unfolds the film scene by scene, in prose, over fifteen to forty pages depending on the production's habits, with barely a few lines of dialogue. In between, many writers go through a step outline, the numbered list of sequences.
These two documents are the most exposed in development. They go round producers and funding panels, and reveal the film's entire mechanism before any contract is signed. Yet copyright, in Switzerland as in most countries, protects a work in its form, not an idea: a three-line pitch can be taken without it counting as copying. If a non-disclosure agreement or an option agreement has already been signed, it often forbids sharing the project with third parties; an AI service that keeps conversations is one.
In IA Confidential, Léa creates a "Writing" space, named after the working title, "Refuge". All the project's conversations will be kept there: one for the pitch, one for the synopsis, one for the treatment.
In the space's Assistant tab, she sets the instructions that will apply to each of these conversations, ahead of her general preferences:
She checks the models. By default, her requests are handled by open models installed on servers in Switzerland and Europe, which keep nothing of her exchanges. For this project, she needs nothing else.
Before writing a single line of the synopsis, Léa wants to know where her idea is fragile. She types her five-line pitch and asks:
I'm developing a feature film. Here is the idea in five lines. Don't write anything: ask me the ten questions a funding panel reader would ask on reading this pitch, about the main character, what she wants, what stands in her way and what the film is really about. Rank them from the most awkward to the least.
The first question stings: why doesn't the warden go down to alert someone as soon as she thinks she recognises the man? The storm answers the "how", not the "why". Léa spends an hour on that one question and finds the answer in the character's past: she has already accused someone wrongly, once, and it cost her the life she had before. The film has just found its subject.
Léa writes the synopsis alone, in one morning. That is where the film's voice is decided, and an AI pulls towards the conventional: its plot suggestions look like films you have already seen. So she uses it as a first reader, not as a pen.
In the afternoon, she pastes her three pages into the synopsis conversation:
Here is my three-page synopsis. Sum up in one sentence what you understood of the story and the ending, then tell me where the narrative skips a step, where the main character's motivation isn't readable, and what you didn't understand. Don't suggest any rewriting.
The one-sentence summary is the most useful test: the assistant understood that the hiker was guilty, whereas the whole film rests on doubt. The synopsis leans too far one way. Léa rewrites the second third to balance the clues. One remark is wrong: the assistant claims that a supporting character disappears without explanation, when their exit is written on page 2. AI can be wrong with complete confidence; Léa checks every remark against her text before touching it.
Going from three pages to twenty means inventing everything the synopsis skips over. Léa first wants to see where the gaps are. It is the request she will make again on every project, and the one to copy:
Based on the attached synopsis, propose a step outline of 50 to 70 sequences for a 100-minute feature in three acts. For each sequence: a number, the location, the time, one sentence on what happens and what changes for the main character. Mark with an asterisk every sequence you had to invent because the synopsis says nothing about it, and list all these gaps together at the end.
Out of fifty-eight sequences, nineteen carry an asterisk. Léa keeps almost none of them: they are plausible and flat. But the list of gaps is valuable. It shows that the second act has only two locations, the common room and the kitchen, and that nothing happens outside between the arrival of the storm and the night of the third day. Léa adds a sequence clearing snow from the roof, where the hiker could let her fall and doesn't. She rewrites the step outline in her own words and files it in the project folder.
The treatment gets written over the day and part of the night. So as not to paste in each successive version again, Léa adds to the space, from her computer with Chrome, the "Refuge" folder, where the notes, the synopsis, the step outline and the treatment in progress live. Adding a folder only works with Chrome, Edge or Opera on a computer; on her tablet, she would simply attach the file to her message. She ticks the "Confidential models" option: the folder will never be sent to an anonymised model.
Signed in to her account, she lets the assistant find the relevant passages in the folder by itself. Every ten pages, she asks it to compare what she has just written with the synopsis and the space instructions: a clue that comes too early, the hut's dog reappearing after being left with someone in the village. She corrects what is right and sets aside what isn't.
The session lasts hours. The context sent to the model is lightened automatically, with the recent exchanges plus a summary of the earlier ones, which keeps answers relevant late in the conversation. Around midnight, she compacts the conversation to start again from a clean base without losing what matters.
The treatment runs to twenty-two pages. Léa asks the assistant to turn it into a document, with a title page and a statement of intent she wrote herself. She rereads it in the preview, then exports it to Word for the producer. The layout is rebuilt from the text: a clean, readable document, not the design of a production package. For the dialogue script, later on, she will keep her usual screenwriting software.
Before sending, she does two things the tool does not do for her. She registers the synopsis and the treatment with a writers' guild or a copyright registration service, to date her work. And she saves a backup of the space to a file on her computer: the history lives only in her browser, and this file will let her restore it without duplicates if she changes machines.
Over these five days, the AI wrote no scene that Léa kept. It asked questions, summarised, counted, compared. That is where it serves a screenwriter: it reads fast, forgets no detail and says bluntly what it did not understand.
Its plot ideas are average by construction. It can invent a detail that is written nowhere, attribute a scene to the wrong act, or miss a subtext. The synopsis and the treatment remain your work, and what they tell is on your name, not the AI's.
With the confidential models, the pitch, the synopsis, the treatment and your exchanges are processed on servers in Switzerland and Europe, under Swiss and European data protection law, outside the reach of the CLOUD Act, and nothing is kept on our side. The conversation history stays in your browser: we could not hand it over if asked, because we do not have it. A folder added with the "Confidential models" option is never sent to an anonymised model.
Anonymised models are a choice you make, never an automatic switch. If you choose one, the Confidentiality filter first checks your message and offers to answer with a confidential model, message unchanged, to send an anonymised version that you review, or to send it as it is. For an unproduced film, bear in mind that a story remains recognisable even when the names have been changed: stay with the confidential models.
Whatever leaves the tool, you are the one taking it out: the exported document, sent to the person you choose, and the backup file, stored wherever you decide. On a shared computer, such as one in a production office, the history can be read by anyone using the same session: work in a personal session and log out when you leave.
When the treatment becomes a script, our guide to writing an unproduced script with AI takes over, sequence by sequence. If the film becomes a series, see how to keep a series bible without it circulating. And before pasting any text into any tool, keep the list of data you should never paste into an AI close at hand.
To put your own pitch to the test of ten questions: open the chat.
