An AI writing companion designed to preserve flow.
About
Most AI writing tools treat writing as a generation problem. They begin with a prompt, produce paragraphs of text, and ask the user to evaluate, edit, or accept what has been written. In this interaction model, the writer gradually shifts from author to reviewer.
Momentum explores an alternative. Rather than generating text, it supports the small moments where writing naturally slows down. The editor waits until the writer hesitates before offering a few nearby continuations that help the text keep moving without taking over the creative process.
The project investigates whether AI can function less as an author and more as a subtle cognitive aid—maintaining momentum while leaving ownership of ideas firmly with the writer. Instead of maximizing AI output, Momentum asks how little intervention is necessary to meaningfully support the act of writing.
Research Question
How can an AI writing tool help a person continue writing without taking over authorship?
A secondary question concerns interaction design rather than language generation:
Can AI become a native part of the writing process instead of existing as a separate conversational interface?
The project is less interested in improving the quality of generated text than in understanding how interface design influences agency, flow, and the feeling of authorship.
Hypotheses
Core hypothesis: writing assistance is most effective when it preserves cognitive momentum rather than replacing cognitive effort. If suggestions are brief, contextual, and appear only during moments of hesitation, writers will maintain a stronger sense of ownership while experiencing fewer interruptions.
Sub-hypotheses:
H1 — Hesitation is the right moment for assistance.
Suggestions triggered after a pause feel less intrusive than continuous autocomplete or proactive generation.
H2 — Small suggestions preserve agency.
Offering short continuations encourages writers to keep thinking instead of evaluating large blocks of generated text.
H3 — Three options are enough.
Presenting exactly three suggestions provides meaningful variation while avoiding unnecessary cognitive load.
H4 — Intent matters more than randomness.
Suggestions representing different intentions—Continue, Clarify, and Turn—support creative thinking more effectively than three equally probable completions.
H5 — Revision should remain local.
Generating alternatives for selected words or short phrases encourages refinement without transforming editing into wholesale rewriting.
H6 — Silence is an interaction state.
Knowing when not to suggest is as important as knowing when to intervene. Suggestions should disappear once the writer resumes control or completes a thought.
Experiments
Momentum serves as a research probe into AI-assisted writing rather than a feature-complete writing application.
Current experiments investigate:
- Whether pause-triggered suggestions feel more natural than continuous autocomplete.
- Whether writers perceive themselves as authors rather than editors of AI-generated text.
- Whether three intention-based suggestions produce more useful interactions than unconstrained generation.
- Whether local revision supports writing flow better than paragraph-level rewriting.
- How latency influences the perception of continuity during writing.
Future iterations will instrument interaction data—including hesitation duration, suggestion acceptance, rejection, regeneration, and abandonment—to study how assistance affects writing behaviour over time.
Alongside implementation, the project is documented through a research journal recording observations, interaction patterns, and design decisions as they emerge throughout development.
Design Decisions
Every interaction in Momentum was designed to minimize interruption while maximizing the writer's sense of agency.
No chat
The interface deliberately removes prompts, conversations, assistant personas, and command-driven interactions.
Rationale:
Writing should remain writing, not a dialogue with software.
Pause-to-suggest
Suggestions appear only after a brief pause in the middle of writing.
Rationale:
Hesitation represents a natural opportunity for assistance instead of constant interruption.
Three suggestion modes
Each suggestion serves a different purpose:
- Continue
- Clarify
- Turn
Rationale:
Different cognitive strategies encourage exploration without increasing interface complexity.
Local assistance
Suggestions complete the current sentence instead of generating entire paragraphs.
Rationale:
AI should extend the writer's thinking rather than replace it.
In-place revision
Selecting a word or phrase generates nearby alternatives.
Rationale:
Revision should feel like sculpting existing text rather than delegating writing to AI.
Keyboard-first interaction
Suggestions can be accepted, skipped, or regenerated without leaving the keyboard.
Rationale:
Maintaining physical flow supports cognitive flow.
Quiet visual language
The interface avoids AI branding, animated assistants, loading states, and visual spectacle.
Rationale:
Attention should remain on the text, not on the technology producing suggestions.
Prototype
Momentum is implemented as a lightweight writing environment built with React, TypeScript, and Tauri, allowing the same codebase to run as both a web prototype and a native macOS application.
Current capabilities include:
- Pause-triggered AI continuations
- Three intention-based suggestions
- Selection-based alternatives
- Rich text editing
- Markdown export
- Local notes
- Support for OpenAI-compatible providers
- Local-first settings
- Native macOS editing commands
The visual language intentionally follows established desktop writing applications, positioning AI as a subtle extension of the editor rather than its primary feature.
Next Research Phase
The current prototype demonstrates an alternative interaction model for AI-assisted writing, but the central research questions remain empirical.
The next stage is a user study investigating:
- whether pause-based suggestions improve perceived writing flow,
- whether users retain a stronger sense of authorship than when using chat-based AI tools,
- how latency influences cognitive momentum,
- how frequently different suggestion types are accepted,
- and whether ambient assistance changes long-term writing behaviour.
Ultimately, Momentum seeks to understand whether AI can become an almost invisible design material – supporting thought without becoming its apparent source.