How the AI assistant works
Use project-aware AI without treating chat as a separate copy of your screenplay.
One assistant, two project contexts
Script and Cinema projects use different writing agents because their creative units are different. The Script assistant works with screenplay nodes, outline, episodes, entities, worldview, knowledge, and shots. The Cinema assistant works with reels, performances, shots, characters, and visual production steps.
In both cases, the assistant does not maintain a hidden second screenplay. It requests browser-side actions, and accepted writes enter the same project systems used by the interface. This is why your role, current episode, selection, and project state matter.
Give every request a landing point
The assistant first has to determine where your request belongs. Make that decision easy by naming one of these:
- the selected text or current scene;
- the outline or a particular beat;
- an episode and scene range;
- a character, location, prop, relationship, or shot;
- a knowledge file or worldview key;
- a reel and production stage in Cinema.
A strong instruction combines target, intention, constraints, and authority: “In episode 3, revise the selected exchange so Nia wins the argument without revealing the key. Keep Sol's final line. Edit the script only.”
Read before broad changes
The assistant can inspect current focus, project entities, episode lists, script ranges, knowledge files, and worldview. Ask for analysis before a risky rewrite. When operating across episodes, it should use actual episode identities and switch context rather than guessing.
Permissions remain real
Viewers receive a read-only tool set. Switching episodes remains possible because it changes only their personal view. Worldview writes need an explicit request in the latest message. Advisors contribute read-only methods; they cannot hire themselves or acquire project permissions.
The assistant's exact internal workflow can change through prompt experiments. Judge it by visible project changes and use bounded requests, selection, history, and review to stay in control.
