SYSTEM DOCS & CHANGELOG
Complete architectural documentation, keyboard shortcodes, local transport protocols, and real-time release history for the Cue AI Prompting Layer.
Getting Started
After installing Cue from the Chrome Web Store or loading unpacked, pin the Cue icon to your browser toolbar. When you visit ChatGPT, Claude, Gemini, or Perplexity, Cue automatically injects a non-intrusive compiler trigger inside the prompt composer.
Keyboard Shortcuts
Trigger prompt optimization without ever taking your fingers off the keys:
Compiler Dial Modes
Normal Mode (Default): Optimizes for conciseness, negative boundaries, tone discipline, and explicit output formats for daily software engineering and executive writing.
Scientific Mode (Pro): Injects formal mathematical proof constraints, adversarial counter-argument verification, KaTeX syntax, and research-grade methodology.
API Key Guides & BYOK Setup
Cue establishes direct HTTPS connections between your browser and your provider with zero intermediate servers:
- Anthropic: Direct console key integration for Claude models. Obtain your API key from console.anthropic.com.
- OpenAI: Paste your
sk-...key from platform.openai.com. - Google Gemini: Get your direct Gemini API key from aistudio.google.com.
- Groq: Ultra-fast LPU inference keys available at console.groq.com.
- DeepSeek: Direct V3 and R1 keys generated from platform.deepseek.com.
- OpenRouter: Multi-model agility with one key across 100+ models from openrouter.ai/keys.
- Custom OpenAI-Compatible Endpoints: Connect any local or self-hosted server using the standard OpenAI schema (e.g., Ollama at
http://localhost:11434/v1, LM Studio athttp://localhost:1234/v1, vLLM, LocalAI, or private company gateways).
Supported Providers (21+ Native & Custom)
Cue features first-class native adapters and catalog routing for 21+ direct foundation model providers plus full support for custom OpenAI-compatible endpoints:
Supported AI Platforms
Cue automatically detects and enhances your prompts directly inside your active chatbox. In Normal / Free mode, Cue works across 15 native built-in AI platforms with zero setup. In Pro mode, you can connect unlimited custom domains and internal tools.
15 Native AI Platforms
15 Native AI Platforms Active (Free & Pro)Cue is natively integrated and automatically active on all 15 portals out of the box:
Unlimited Custom Domains & Private Company Tools
Need Cue on an internal corporate portal, staging environment, or self-hosted LLM UI? Cue Pro unlocks universal domain injection. Simply open the Cue extension settings or dashboard, enter any domain (e.g. ai.company.internal, openwebui.myteam.com, or localhost:3000), and Cue will seamlessly mount directly into that application's composer.
Zero-Telemetry Architecture
Your prompt data and private thoughts are yours alone. Cue enforces absolute local boundary integrity:
- Local-First Storage: Keys and history remain stored inside your browser's encrypted storage.
- Zero Proxy Relays: HTTPS requests travel directly from your client to Anthropic, OpenAI, or Google.
- Encrypted Vault Protection: Optional PBKDF2/AES-GCM key locking with memory wiping on idle.
Research & Empirical Claims
Cue's efficiency metrics (time saved, turns reduced, and context tokens conserved) are grounded in published empirical research across transformer self-attention mechanisms, reasoning models, and prompt steering.
When prompting raw LLMs casually, conversational drift forces 3 to 5 corrective turns ("no, don't do that", "make it cleaner", "you forgot the types"). Each turn resubmits preceding context into the transformer window, quadratic in cost and high in human cognitive fatigue. Cue compiles the full turn-one specification before execution.
Demonstrated that conditioning LLMs with specialized practitioner personas steers the neural network into dense latent clusters, dramatically reducing hallucination rates and increasing domain specificity compared to raw zero-shot directives.
Proved that structured multi-step reasoning breakdowns overcome autoregressive attention bottlenecks, boosting multi-step logic and mathematical accuracy by 200% to 300% over direct response prompting.
Showed that decomposing a complex query into sub-problems (Context → Objectives → Edge Cases → Deliverable Schema) enables frontier models to solve compositional tasks that standard single-step prompts consistently fail.
Documented that delineating user input and system rules with structural XML tags prevents prompt injection, instruction bleeding, and ambiguous parsing across Claude and next-generation reasoning architectures.
Demonstrated that language models struggle to parse pure negative constraints ("do not do X"). Peak reliability is achieved only by pairing negative prohibitions with immediate positive replacement contracts ("Do not provide conversational filler; output the raw executable payload directly").
Troubleshooting & Feedback
Experiencing an issue or have a feature request? Here are the most common diagnostic resolutions: