OS1

From a scene to something you can use.

What would it take to make the computer in Her work outside the film?

OS1 starts with the relationship between a person, a conversation, and the work on their screen. The first scene to explore is letter writing: speaking a thought, seeing it appear on paper, then asking for a change without leaving the conversation. It gives the project something concrete to make real.

A film can cut away before an awkward pause or a lost connection. A working version has to keep going. It needs to know which letter is open, distinguish a sentence from an instruction, and keep the words when the voice stops. Recreating the scene turns each of those invisible details into a design problem.

This is an exploration of that working relationship. The project uses its own available voices; it does not recreate Samantha’s identity or voice.

Let the work occupy the screen.

Geoff McFetridge’s accounts of designing Her describe powerful technology expressed through familiar, unobtrusive forms. His discussion of color, a painting’s frame, and the space around an interface became useful references for this project. [1]

The installation begins with a coral field and white strands that gather into a ring. An amber horizon gives the screen depth. Around it, a white mat and a small exhibition label treat the running software as an artwork. Conversation carries the controls; the page can give its attention to the letter.

Letter Writer keeps its own paper, photographs, and task context. OS1 supplies a voice relationship with that application. Asking to write brings the app forward while the conversation continues. This follows a key project boundary: an intelligent system can work with an application without replacing its interface. The film’s handwritten-letter screens were an important visual reference. [2]

Underneath the conversation.

The browser owns the microphone, audio playback, and the visible scene. A persistent agent, built on OpenAI’s open-source Codex runtime, works with the application. The current voice layer uses GPT-Live, with an independent transcription path for dictation. [3] [4]

OS1

Keeps the conversation, media lifecycle, permissions, and active scene together.

Agent

Interprets a request and proposes an action using the context of the current task.

App

Owns the letter and its state. It validates the action, saves the result, and renders it.

For example, a spoken request to revise a sentence becomes a structured edit: a task ID, the revision being edited, and the proposed new text. The app validates that request before changing the letter. A project-owned Skill describes how the agent should collaborate with the app; it does not grant unrestricted access to the page or the user’s data.

This separation makes a quiet interface possible. The agent can ask for a meaningful change, and the software can decide whether that change is valid. No pointer coordinates or arbitrary page manipulation are needed for the writing flow.

The words should stay.

Dictation and instruction travel through different paths. While a letter is being dictated, stable transcription goes directly into its working text. It does not wait for the agent to finish a conversational turn. Requests addressed to OS1 are handled as control speech rather than letter content. [5]

Unfinished but stable text is saved as a recoverable prefix. Pausing, leaving the app, or losing a connection promotes that prefix into the draft. A refresh restores the work in a paused state, so an ordinary conversation does not accidentally become the next sentence of the letter.

Handwriting has a similar continuity problem. Adding a word should extend the line, without redrawing the ink that is already there. The handwriting renderer carries its stroke-generation state forward as text arrives, building on Alex Graves’s sequence-generation work and the Longhand implementation. [6] [7]

The browser experience uses the same handwriting engine as the local version. New words extend the existing pen strokes, so the deployed work keeps the same hand and writing motion.

A few small decisions.

Installation is already part of the experience.

After the entry confirmation, installation starts by itself. Prerecorded questions give its opening a consistent rhythm. Answers are treated as limited evidence; a missed answer must not trap someone in setup or become an invented personality profile.

One voice relationship, even across tabs.

Each active user has an independent runtime. Only one browser page owns its live microphone connection at a time. A new page can take over the same workspace, while the old one releases its audio. The letter and the original experience deadline remain.

The edge of the frame has a job.

A compact Visit time label sits in the upper-left mat, diagonally opposite the artwork label. It makes the remaining time visible without taking over the conversation. Entry notes, actions, and the timer share fixed positions, so the surrounding frame stays still as the experience changes.

The current prototype focuses on conversation, dictation, and spoken editing in Letter Writer. Sending and printing are not implemented. These notes will grow with the software and the recordings.