Spell

Spell integrates language models into the document editing interface to provide contextual writing suggestions, completions, and rewrites as users type. The system likely uses token streaming and latency-optimized inference to surface suggestions without blocking the editing experience, with a suggestion UI overlay that allows accept/reject/modify workflows similar to GitHub Copilot but adapted for prose rather than code.

Spell provides targeted rewriting capabilities that transform selected text or entire sections into different tones, styles, or formats (e.g., formal to casual, passive to active, summary to expansion). This likely uses prompt engineering with style classifiers or fine-tuned models to maintain semantic meaning while shifting linguistic properties, with the rewritten text presented as alternatives the user can accept or iterate on.

Spell supports real-time multi-user editing with conflict resolution that may leverage AI to intelligently merge concurrent edits. When multiple users edit the same section, the system likely uses operational transformation (OT) or CRDT-based merging, with AI potentially assisting in resolving conflicts by understanding semantic intent rather than just applying last-write-wins or manual merge strategies.

Spell analyzes document content to extract or generate hierarchical outlines, section summaries, and structural metadata. This likely uses NLP techniques (entity recognition, semantic segmentation, or transformer-based section detection) to identify document sections, headings, and logical flow, then generates or refines outlines that can be used for navigation, reorganization, or content planning.

Spell integrates research capabilities that allow users to cite sources, pull in external content, or generate citations in standard formats (APA, MLA, Chicago, etc.). This likely involves API integrations with citation databases or web search APIs, combined with prompt engineering to format citations correctly and embed source references directly into the document with proper attribution.

Spell allows users to generate content from scratch using AI by providing prompts or selecting from predefined templates. This likely uses a prompt engineering system with template variables (e.g., 'Generate a [TONE] email to [RECIPIENT] about [TOPIC]') that are filled in by the user, then passed to an LLM for generation. Generated content is inserted directly into the document with options to regenerate, edit, or discard.

Spell provides real-time or on-demand grammar, style, and clarity checking that goes beyond simple rule-based linting. This likely uses transformer-based models or fine-tuned classifiers to detect issues like awkward phrasing, unclear pronoun references, passive voice overuse, or readability problems, with suggestions for improvement that maintain the original intent while improving clarity.

Spell can generate summaries of document content at various levels of detail (executive summary, bullet points, one-paragraph summary, etc.). This likely uses abstractive summarization models (transformer-based, such as BART or T5) that generate new text capturing the main ideas, rather than extractive methods that just pull existing sentences. Summaries can be generated on-demand or automatically updated as the document changes.