Changelog

Output language selector supporting 10 languages: English, Spanish, French, German, Russian, Ukrainian, Armenian, Chinese, Japanese, and Korean. The selected language is injected into the system prompt as a primary language directive.

Domain/industry field in Persona & Targeting section. The engine uses this to select domain-specific terminology and adjust framework weighting (e.g., TRACE for technical domains, SCOPE for business).

Complexity level selector (simple, moderate, advanced, expert) that maps to technical depth modifiers in the generated prompt. Expert-level prompts include schema definitions, validation criteria, and multi-step reasoning chains.

Optimization priority selector (speed, balanced, quality, creative) that influences framework selection heuristics. Speed priority favors RTF/APE/TAG; quality priority favors CO-STAR/RISE/TRACE.

Framework selection logic rewritten with a 4-step analysis pipeline: (1) read all user fields, (2) identify structural needs, (3) match to optimal framework, (4) annotate selection rationale. Eliminates CO-STAR default bias.

System prompt now enforces framework-specific section headers in output. RTF outputs use ### ROLE / ### TASK / ### FORMAT; PAIN outputs use ### PROBLEM / ### ACTION / ### INFORMATION / ### NEXT STEPS, etc.

Complete prompt history system with per-generation snapshots. Each entry stores all input fields (goal, background, role, audience, format, tone, examples, constraints), selected model, and the full generated output.

One-click history restore: selecting a previous generation repopulates all form fields and model selection to their exact state at generation time.

Per-item history deletion and bulk clear with organization-level data isolation. History is filtered by Clerk orgId — switching organizations shows only that org's generations.

Output terminal with real-time character, word, and line count. Includes copy-to-clipboard, download-as-Markdown (.md), and auto-scroll-to-output on generation complete.

Input validation system with real-time character counting on goal and background fields. Goal requires minimum 10 characters. Visual indicators shift from red → yellow → green as input quality improves.

Dashboard restructured into 4 collapsible sections: Core Objectives, Persona & Targeting, Output Configuration, and Advanced Configuration. Reduces cognitive load for simple tasks while exposing full control for power users.

Claude Sonnet 4.5 (Anthropic) added to model selector. 200K context window, highest reasoning capability for complex prompt engineering and nuanced multi-step tasks.

Grok 4.1 Fast (xAI) added as an experimental model with real-time knowledge access and 128K context. Marked as Beta in the model selector.

Gemini 3 Flash Preview (Google) added with 1M token context window and sub-second latency. Optimized for rapid iteration workflows.

Model configuration modal rebuilt with detailed spec cards showing context window size, speed rating, and cost tier for each model. Includes provider attribution and selection tips.

Model selection state now persists to localStorage (key: advi_selected_model) and correctly syncs between the settings modal and API request payload without false 'unsaved changes' indicators.

Deprecated model entries (GPT-3.5 Turbo, older Gemini variants) removed from the selector. All traffic consolidated to 5 actively maintained model endpoints via OpenRouter.

Initial framework selection engine supporting 20 prompt engineering methodologies: CO-STAR, CREATE, PROMPT, Chain of Thought, Self-Consistency, RISE, SCOPE, STAR, ROSES, TRACE, 5W1H, CARE, GRADE, RTF, APE, ERA, TAG, BAB, SOAR, PREP, PAIN, and Few-Shot.

Constraint injection system with MUST/ALWAYS positive directives and NEVER/AVOID negative directives. Constraints are placed at both the beginning and end of the generated prompt for reinforcement.

Persona engineering layer in the system prompt. Transforms generic role assignments (e.g., 'marketing expert') into deep specifications with credentials, analytical frameworks, and behavioral constraints.

Few-Shot example injection field. User-provided input-output patterns are formatted as exemplars in the generated prompt, enabling direct behavioral conditioning through demonstration.

System instruction expanded to a 3-phase pipeline: (1) Requirements Analysis — extracts intent, assesses complexity, maps constraints; (2) Framework Selection — matches inputs to the optimal methodology; (3) Advanced Optimizations — applies structural delimiters, cognitive load optimization, and output determinism.

POST /api/refine endpoint: accepts goal, background, role, audience, format, tone, examples, constraints, and model fields. Returns a single refined prompt string optimized for frontier LLMs.

Clerk authentication integration with organization-level access control. API requires valid userId and orgId — requests without org membership return 401.

OpenRouter API integration as the unified model gateway. All model requests route through OpenRouter with configurable API key, site URL tracking (HTTP-Referer), and app name headers (X-Title).

Error handling pipeline with typed responses: 400 (invalid input), 401 (unauthorized), 429 (rate limited), 503 (provider config error), 504 (timeout). All errors include human-readable messages.

API temperature set to 0.7 with 4,000 max tokens. Balances creative variation with output consistency for prompt engineering workloads.

Prompt refinement workspace with structured input fields: goal (required), background, role, audience, format, tone, examples, and constraints.

Default model set to meta-llama/llama-3.1-8b-instruct:free via OpenRouter. 128K context window, zero cost, suitable for general-purpose prompt generation.

GPT-4o (OpenAI) available as a premium model option. 128K context, fast inference, consistent performance across task types.

Real-time prompt generation with streaming output display. Generated prompts are rendered in a terminal-style output pane with syntax-aware formatting.

Clipboard copy functionality for generated prompts. Single-click copies the full output with a visual confirmation state (2-second timeout).