LoomScribe is a premium, highly specialized creative writing workspace designed specifically for DeepSeek models. It provides a dual-slot compiler architecture that maximizes API cache performance while putting granular writing controls (POV, sensory intensity, dialogue register, character pushback, complications, outline sandbox mode, and premises generation mode) directly in your hands.
Built with a lightweight Node.js/Express backend and a responsive, zero-build vanilla HTML/CSS/JS frontend, LoomScribe serves as a distraction-free desktop environment for drafting, brainstorming, and editing interactive fiction.
DeepSeek models utilize a key-value (KV) cache of the prefix prompt. To prevent cache-busting, LoomScribe compiles your prompt into two distinct slots:
- Slot 1 (System Prompt - Stable): Contains the foundational character, prose, tone, and formatting rules. This stays byte-for-byte identical across turns to maximize cache hits, reducing cost and latency.
- Slot 2 (Post-History - Dynamic): Injected after all chat history, immediately before the model generates. This holds dynamic turn-based instructions such as word count targets, pushback, complications, premises-generation directives, and the Director's Note. Changes here never bust the cache.
Loaded directly from a central JSON schema, the right pane provides a live suite of parameters:
- System-Slot Controls (Busts Cache - Amber Warning Dot):
- Point of View: Toggle between Close Third, Deep First, and Omniscient POV blocks.
- Scene Intensity: Adjust from Tender, Sensory, Charged, to Raw narrative tone and description depth.
- Dialogue Style: Control dialogue styling (Silent, Playful, Candid, Commanding).
- POV Focus Spotlight: Shift narrative focus between Balanced, POV Interiority, and Partner Reaction.
- Outline / Brainstorm Mode: Switch the model into planning mode instead of full prose.
- Premises & Ideas Mode: Switch the model into six-premise generation mode instead of prose.
- Post-History Controls (Cache Safe):
- Response Length: Slider targeting a word count between 600 and 3,000 words.
- Pushback / Resistance: Graded slider (1 to 5) controlling how cooperative the AI-controlled characters are (Compliant, Realistic, Resistant).
- Complication Generator: Toggle to introduce sudden external distractions, guilt, hesitations, or obstacles.
- Director's Note: Per-turn free text (e.g.,
"Focus on the rain sound; build slow tension") appended at high recency.
- Advanced Block Overrides: Manually override individual system-prompt markdown files inside your preset.
- Live Prompt Compiler Preview: View compiled Slot 1 and Slot 2 prompt states in real-time.
A built-in preset manager interface allows you to create, edit, duplicate, delete, and import narrative presets directly.
- Two-Column Editor: Modify system body prompts, metadata, categories, and parameter overrides.
- File Dropzone: Drag and drop JSON preset files to import them instantly.
- Server CRUD Endpoints: Fully backed by API routes for managing local presets on the fly.
A dedicated toggle shifts LoomScribe from prose drafting into a planning sandbox. Enabling Outline Mode disables the prose-focused blocks, then injects directives that ask the model to expand on plot structure, scene beats, and character trajectories instead of writing finished chapters.
A separate toggle switches the engine into a six-premise brainstorming mode. When enabled, the compiler disables prose-focused blocks and appends a directive that asks the model to generate exactly six fully developed premises, each with a distinct tonal lane and a clear conflict axis. This is for ideation, not chapter writing.
Switch between chat threads while other responses stream in the background. Generating threads show a pulsing ⚡ status dot in the sidebar and rebuild cleanly when revisited. Streaming now runs over WebSockets, which improves reconnects, background updates, and per-thread aborts.
Each message has quick actions for editing, regenerating, and copying raw content to the clipboard. User input drafts are preserved per conversation, so switching threads or refreshing the app does not wipe your draft.
LoomScribe tracks conversation history as a branching version tree. Inline editing of any user message branches a new lineage, and prev/next controls let you move between alternative timelines, drafts, and regenerations.
A server-side logger format writes HTTP requests, WebSocket connection updates, and prompt compilation details as structured JSON events. This standardizes local application logging and eases troubleshooting.
Make sure you have Node.js installed (LoomScribe works on any recent LTS version).
Clone the repository and install the dependencies:
npm installCopy the template environment file:
cp .env.example .envOpen .env and configure your settings:
PORT: Port to host the app (defaults to3000).DEEPSEEK_API_KEY: Set yoursk-...API key here. (Alternatively, configure it securely in the UI sidebar settings).APP_PASSWORD: Set a password to protect the LoomScribe workspace interface. (If left blank, authentication check is skipped).
Run the development command:
npm run devOpen http://localhost:3000 in your browser.
On Windows: Simply double-click start-loomscribe.bat. It will start the server, open the browser automatically, and cleanly shut down the background processes when you press any key in the console window.
loomscribe/
├── data/
│ └── db.json # Server-side JSON database (settings, chats, messages)
├── engine/
│ ├── blocks/ # Reusable Markdown prompt blocks
│ │ ├── index.json # Block registry (id, file, group, order)
│ │ ├── base_writer.md # Core AI identity
│ │ ├── tone_register.md # Baseline prose standards
│ │ ├── format_rules.md # Layout, markdown formatting, paragraph structures
│ │ ├── premises_mode.md # Instructions for six-premise generation
│ │ ├── pov_*.md # Point-of-view definitions (first, third, author)
│ │ ├── focus_*.md # POV spotlight controls (self, partner, balanced)
│ │ ├── intensity_*.md # Scene intensity levels (tender, sensory, charged, raw)
│ │ └── dialogue_*.md # Dialogue styles (silent, playful, candid, commanding)
│ ├── presets/ # Preset JSON files (defining categories & defaults)
│ ├── compiler.js # Resolves schema parameters and compiles dual-slot output
│ ├── schema.json # Parameter definitions loaded by the UI to render inputs
├── public/ # Zero-build frontend files
│ ├── css/ # Modular stylesheet components (layout, chat, sidebar, right-pane)
│ ├── js/ # Domain modules (api, ui, auth, chat, version-tree, preset-manager)
│ ├── index.html # Main application entry point
│ └── login.html # Authentication login screen
├── src/server/ # Express backend source code
│ ├── endpoints/ # Modular REST endpoint routes (conversations, config, presets, auth)
│ └── routes.js # API router integration
└── server.js # Application main runner file
When compiling a turn's prompt, engine/compiler.js automatically groups instructions to maximize caching:
- Block Selection: The compiler selects the appropriate POV, scene intensity, dialogue style, and focus spotlight blocks based on settings.
- Concat Slot 1: Integrates the stable preset body and foundational blocks.
- Concat Slot 2: Integrates the dynamic length targets, character pushback settings, complications, and the user's Director's Note.
- Amber dot next to settings: Indicates that changing this slider/select controls Slot 1 blocks and will invalidate the KV cache on the next token request.
- Green/Safe controls: Per-turn parameters and Director's Notes belong in Slot 2. They can change every turn without busting the KV cache.
A preset is a single JSON file dropped in engine/presets/. It will load into the preset picker automatically.
See PROMPT_ENGINE.md for instructions on creating new presets. You can use the provided PRESET_CREATOR.md prompt with any LLM to automatically generate a preset for any scenario.
Use outline_mode for brainstorming/plotting presets and premises_mode for presets that should generate exactly six fully developed premises instead of chapter prose.
- PROMPT_ENGINE.md: Deep dive into the prompt compiler, block registry, and custom presets mapping.