From Draft to Done: Tensor's Writing Assistant

You write dozens of messages every day. Emails, Slack messages, form responses, social media posts, document comments, pull request descriptions. Each one has a different audience, a different tone, a different set of expectations. And yet most AI writing tools treat all text the same: paste your draft, get a generic revision, hope it sounds like you. Tensor's writing assistant works differently. It lives inside every text field in your browser, understands who you are writing to and why, and produces text that sounds like you on your best day.

Inline Everywhere

The writing assistant is not a separate app or a popup window. It activates inside any text field you type in: Gmail compose windows, Slack message boxes, Notion pages, Google Docs, GitHub issue descriptions, LinkedIn posts, and any standard text input or contenteditable element. When you start typing, Tensor monitors your input and offers assistance through a subtle floating toolbar that appears near your cursor.

The toolbar offers four core actions:

• Fix — correct grammar, spelling, and punctuation without changing your voice
• Improve — enhance clarity, conciseness, and flow while preserving your intent
• Adjust tone — shift between formal, casual, friendly, professional, assertive, or empathetic
• Compose — generate a full draft from a brief description of what you want to say

Each action can be triggered with a keyboard shortcut or a single click. The result appears as a diff overlay on your existing text, showing exactly what changed. You can accept all changes, reject them, or cherry-pick individual edits. Nothing is ever modified without your explicit approval.

Grammar That Understands Context

Grammar checking sounds like a solved problem, but most tools still get it wrong in important ways. They flag intentional fragments in casual messages. They suggest overly formal alternatives in friendly emails. They do not understand that different contexts have different rules. An academic paper has different grammar expectations than a Slack message to your team.

Tensor's grammar engine understands context at multiple levels. It considers the platform you are writing on, whether that is Gmail, Slack, Twitter, or a Google Doc. It considers the formality level based on your recipient, pulling from your Personal Context data if available. It considers the genre, distinguishing between a bug report, a thank-you note, a project proposal, and a casual update.

// Context-aware grammar adjustment
{
  platform: "slack",
  channel: "#engineering",
  formality: "casual-professional",
  rules: {
    fragments: "allow",           // "Looks good. Ship it." is fine
    contractions: "prefer",       // "don't" over "do not"
    sentenceCase: "flexible",     // lowercase starts are okay
    emojiSuggestions: true,       // suggest relevant emoji
    jargon: "allow-technical"     // domain terms are not flagged
  }
}

This context awareness means the assistant never suggests replacing "lgtm" with "This looks great to me" in a code review. It knows that abbreviations and shorthand are normal in that context. But if you are writing a client-facing email, it will gently suggest expanding acronyms and using complete sentences.

Tone Adjustment: More Than a Slider

Tone is subtle. The difference between assertive and aggressive, between casual and unprofessional, between empathetic and patronizing, often comes down to a few word choices. Tensor's tone adjustment does not simply swap vocabulary. It restructures sentences, adjusts pacing, modifies hedging language, and recalibrates the emotional register of your text.

Consider a simple feedback message. The same core message can be expressed in radically different tones. A direct tone might open with the issue immediately and state what needs to change. A diplomatic tone might lead with acknowledgment of what went well before raising the concern. An empathetic tone might start by validating the difficulty of the task before offering suggestions.

The tone engine models six dimensions of communication style: formality (casual to ceremonial), directness (hedged to blunt), warmth (detached to effusive), authority (deferential to commanding), complexity (simple to sophisticated), and energy (reserved to enthusiastic). When you select a tone preset, it adjusts all six dimensions simultaneously. You can also fine-tune individual dimensions for precise control.

Personal Context Integration

The most powerful feature of Tensor's writing assistant is its integration with Personal Context, the on-device knowledge base that stores information about you, your relationships, your projects, and your communication preferences. When you compose an email to a specific person, the assistant can reference your history with them: the tone you typically use, topics you have discussed, follow-ups you promised, and even their communication preferences if you have noted them.

For example, if you are replying to an email from a colleague named Sarah about a project deadline, the assistant knows that Sarah prefers concise emails, that you promised her an update by Friday, that the project name is "Atlas Migration," and that the current status is blocked by a dependency on the infrastructure team. With this context, it can draft a reply that is specific, relevant, and sounds like you wrote it, because in a sense you did: the assistant is drawing on your own words and preferences.

Personal Context is stored entirely on your device. No email content, relationship data, or communication history is sent to any server. The writing model runs locally, and the Personal Context data never leaves your machine.

Email Composition: From Bullet Points to Polished Drafts

The compose feature transforms brief notes into complete emails. You type a few bullet points describing what you want to communicate, select a tone and formality level, and the assistant generates a full draft. But unlike generic AI email writers, Tensor's compose feature draws on your actual writing style.

During your first week of use, the writing assistant analyzes your sent emails (with your permission and entirely on-device) to build a style profile. It learns your typical greeting and sign-off patterns, your average sentence length, your vocabulary preferences, your use of humor or formality, and your structural tendencies. Some people write short paragraphs with direct sentences. Others write flowing paragraphs with subordinate clauses. The style profile captures these patterns so that generated text blends seamlessly with your organic writing.

// Style profile (simplified)
{
  greeting: "Hey {firstName},",
  signoff: "Best,\nYiming",
  avgSentenceLength: 14.2,
  paragraphLength: "short",    // 2-3 sentences
  contractions: 0.82,          // uses contractions 82% of the time
  passiveVoice: 0.08,          // rarely uses passive voice
  hedgingFrequency: "low",     // direct communication style
  questionRate: 0.15,          // ends ~15% of messages with a question
  emojiUsage: "occasional"
}

Multi-Language Support

The writing assistant supports composition and editing in over 30 languages, with grammar checking available in 12 languages. More importantly, it supports seamless bilingual workflows. If you receive an email in German and want to reply in English, or if you need to draft a message in Spanish based on English notes, the assistant handles the translation while maintaining your personal style.

The translation is not a separate step. You type in whatever language is natural for your thinking, mark the sections you want translated, and the assistant produces output in the target language with appropriate cultural adjustments. Formal register in Japanese is very different from formal register in English, and the assistant accounts for these cultural norms.

Smart Suggestions and Autocomplete

Beyond full composition, the assistant offers real-time autocomplete suggestions as you type. These are not generic text predictions. They are context-aware completions that consider what you are writing about, who you are writing to, and your personal style. When you start typing a project update email, the autocomplete draws on your recent project data to suggest relevant details and metrics.

Suggestions appear as ghosted text ahead of your cursor, similar to GitHub Copilot but for natural language. Press Tab to accept a suggestion, or keep typing to dismiss it. The suggestion engine runs locally and responds in under 50 milliseconds, fast enough to feel like a natural extension of your typing rather than an interruption.

The writing assistant is designed around one principle: your words should sound like you, only better. It does not impose a house style or a corporate voice. It amplifies your natural communication patterns while catching the errors, inconsistencies, and rough edges that everyone produces when writing quickly. The result is that you spend less time editing and more time communicating.