Core Principles

Fundamental principles guiding Doxen.ai product design

1. User Agency: Copilot and Captain

The user is the sole decision-maker. AI is your capable copilot, but you are always the captain.

This is Doxen.ai's most fundamental design philosophy. All other principles serve this one:

┌─────────────────────────────────────────────────────────────────┐
│                                                                 │
│                      User (Captain)                             │
│                           │                                     │
│              ┌────────────┼────────────┐                        │
│              │            │            │                        │
│              ▼            ▼            ▼                        │
│         ┌────────┐  ┌──────────┐  ┌─────────┐                  │
│         │ Decide │  │ Confirm  │  │ Control │                  │
│         └────┬───┘  └────┬─────┘  └────┬────┘                  │
│              │           │             │                        │
│              └───────────┼─────────────┘                        │
│                          │                                      │
│                          ▼                                      │
│                    ┌──────────┐                                 │
│                    │    AI    │                                 │
│                    │ Copilot  │                                 │
│                    └──────────┘                                 │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Design Patterns

Pattern	Description
Explicit Confirmation	No AI-generated content is finalized without a user action (Tab to accept, Click to apply)
Full Reversibility	Every AI intervention is logged and can be undone with a single action

Concrete Manifestations

AI Never Acts Unilaterally
- All modifications require user confirmation to take effect
- Even the lightest completion needs Tab to confirm
- Generated content requires user decision to commit
User Can Always Intervene
- Tasks can be paused, cancelled, modified
- Execution mode can be switched at any time
- Parameters can be adjusted at any time
Complete Transparency of History
- All AI operations are recorded
- Any modification can be reverted
- Rejected suggestions are also preserved for review

2. Unobtrusive Assistance

AI should behave as a "Silent Partner," emerging only when needed and respecting the user's flow.

We avoid the "Clippy" problem of being overly chatty or intrusive.

Mechanisms

Mechanism	Description
Passive Indicators	AI suggestions appear as subtle visual cues (e.g., a small icon in the gutter) rather than popping up dialogs
Heuristic Triggers	Proactive suggestions are only shown when signals (user pause, cursor at sentence boundary) indicate a good intervention moment

Principle: Better to disturb less than make users feel harassed.

3. Interaction-as-Prompt

User operations themselves are "prompts" — reducing the Token Cost of Interaction.

Instead of requiring a textual prompt ("Summarize this paragraph"), the system infers intent from interaction:

User Action	AI Understands Intent
Select a paragraph + click "Simplify"	Simplify this paragraph, preserve meaning
Pause typing for 3 seconds	May need completion suggestion
Select text + invoke "Refine" shortcut	Polish this selection in context

This reduces the cognitive effort needed to describe intent in words.

4. Context Awareness (Hierarchical Context)

AI requires context. Doxen maintains Full State Awareness through a hierarchical context window.

Layer	Scope	Refresh Rate
Local Context	Cursor, paragraph, recent edits	High
Global Context	Document structure, headings	Medium
Workspace Context	Cross-document knowledge, glossary	Via RAG

AI should "understand" what user is doing, reducing the burden of explaining background.

5. Non-Destructive Design (Proposal-Based Mutation)

AI acts as a layer of assistance and cannot mutate document state directly. All AI outputs are Proposals.

Adaptive Representation

The UI dynamically chooses the best presentation form based on the proposal's nature:

Proposal Type	Presentation
Completion (L0)	AI Completion: Gray inline suggestions
Minor correction (L1)	Diff View: Inline annotation bubble
Paragraph change (L2)	Comment Bubble: Side panel annotation
Structural change (L3-L4)	Generative Block: Review panel

Commit Flow

AI generates content → Proposal (immutable) → User reviews → Commit or Reject → History recorded

6. Traceability

Every document change can be traced to its source, supporting accountability and quality review.

Every AI operation is a "Commit":

Document: business-plan.md
├── [v1] Initial draft                    ← manual save
├── [v2] Task: Polish summary (committed)  ← AI task
├── [v3] Task: Expand analysis (committed) ← AI task
├── [v4] Manual edit                       ← user edit
└── [v5] Task: Polish conclusion (rejected) ← rejected task

Implementation

Attribution: Metadata tagging (source: user vs source: ai)
Version History: Atomic commits for accepted proposals
Revert Has No Cost: Can revert to any historical version at any time

7. Quality over Quantity

We prioritize precise, high-quality suggestions that improve the document's value over generating large volumes of generic text.

Our goal is not to simply generate text for the user, but to help them write better documents.

Summary

These principles have one core goal:

Make AI Useful: Cover the entire writing lifecycle, truly improve efficiency
Make AI Controllable: Users always feel in control, never feel out of control
Make AI Trustworthy: Predictable behavior, traceable results, revertible errors
Make AI Easy to Use: No need to learn prompts, interaction is the instruction

Remember: Doxen.ai's goal is not to have AI write for users, but to help users complete their own writing more efficiently.

Core Principles

On this page