feat: v1.3.0 — plan-first workflow, OpenRouter provider, enhanced prompt engine

Major changes: - Plan-first workflow: AI generates structured plan before code, with plan review card (Modify Plan / Start Coding / Skip to Code) - Post-coding UX: Preview + Request Modifications buttons after code gen - OpenRouter integration: 4th AI provider with 20+ model support - Enhanced prompt engine: 9 strategies, 11+ intent patterns, modular - PLAN MODE system prompt block in all 4 services - Fixed stale React closure in approveAndGenerate with isApproval flag - Fixed canvas auto-opening during plan phase with wasIdle gate - Updated README, CHANGELOG, .env.example, version bump to 1.3.0
2026-03-18 18:45:37 +00:00
parent cca11fe07a
commit a4b7a0d9e4
17 changed files with 3189 additions and 358 deletions
--- a/lib/services/adapter-instance.ts
+++ b/lib/services/adapter-instance.ts
@@ -1,4 +1,5 @@
 import ModelAdapter from "./model-adapter";
+import { OpenRouterService } from "./openrouter";

 const adapter = new ModelAdapter();

--- a/lib/services/model-adapter.ts
+++ b/lib/services/model-adapter.ts
@@ -2,6 +2,7 @@ import type { ModelProvider, APIResponse, ChatMessage, AIAssistMessage } from "@
 import OllamaCloudService from "./ollama-cloud";
 import ZaiPlanService from "./zai-plan";
 import qwenOAuthService, { QwenOAuthConfig, QwenOAuthToken } from "./qwen-oauth";
+import { OpenRouterService } from "./openrouter";

 export interface ModelAdapterConfig {
  qwen?: QwenOAuthConfig;
@@ -14,17 +15,22 @@ export interface ModelAdapterConfig {
    generalEndpoint?: string;
    codingEndpoint?: string;
  };
+  openrouter?: {
+    apiKey?: string;
+  };
 }

 export class ModelAdapter {
  private ollamaService: OllamaCloudService;
  private zaiService: ZaiPlanService;
  private qwenService = qwenOAuthService;
+  private openRouterService: OpenRouterService;
  private preferredProvider: ModelProvider;

  constructor(config: ModelAdapterConfig = {}, preferredProvider: ModelProvider = "ollama") {
    this.ollamaService = new OllamaCloudService(config.ollama);
    this.zaiService = new ZaiPlanService(config.zai);
+    this.openRouterService = new OpenRouterService(config.openrouter);
    this.preferredProvider = preferredProvider;

    if (config.qwen) {
@@ -62,6 +68,10 @@ export class ModelAdapter {
    this.qwenService.setOAuthTokens(tokens);
  }

+  updateOpenRouterApiKey(apiKey: string): void {
+    this.openRouterService = new OpenRouterService({ apiKey });
+  }
+
  async startQwenOAuth(): Promise<QwenOAuthToken> {
    return await this.qwenService.signIn();
  }
@@ -90,6 +100,8 @@ export class ModelAdapter {
        return this.ollamaService.hasAuth();
      case "zai":
        return this.zaiService.hasAuth();
+      case "openrouter":
+        return this.openRouterService.hasAuth();
      default:
        return false;
    }
@@ -114,6 +126,8 @@ export class ModelAdapter {
        return this.ollamaService;
      case "zai":
        return this.zaiService;
+      case "openrouter":
+        return this.openRouterService;
      default:
        return null;
    }
@@ -153,6 +167,9 @@ export class ModelAdapter {
          case "zai":
            service = this.zaiService;
            break;
+          case "openrouter":
+            service = this.openRouterService;
+            break;
        }

        const result = await operation(service);
@@ -183,26 +200,26 @@ export class ModelAdapter {
    };
  }

-  async enhancePrompt(prompt: string, provider?: ModelProvider, model?: string): Promise<APIResponse<string>> {
-    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai");
+  async enhancePrompt(prompt: string, provider?: ModelProvider, model?: string, options?: { toolCategory?: string; template?: string; diagnostics?: string }): Promise<APIResponse<string>> {
+    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai", "openrouter");
    const providers: ModelProvider[] = provider ? [provider] : fallback;
-    return this.callWithFallback((service) => service.enhancePrompt(prompt, model), providers);
+    return this.callWithFallback((service) => service.enhancePrompt(prompt, model, options), providers);
  }

  async generatePRD(idea: string, provider?: ModelProvider, model?: string): Promise<APIResponse<string>> {
-    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai");
+    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai", "openrouter");
    const providers: ModelProvider[] = provider ? [provider] : fallback;
    return this.callWithFallback((service) => service.generatePRD(idea, model), providers);
  }

  async generateActionPlan(prd: string, provider?: ModelProvider, model?: string): Promise<APIResponse<string>> {
-    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai");
+    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai", "openrouter");
    const providers: ModelProvider[] = provider ? [provider] : fallback;
    return this.callWithFallback((service) => service.generateActionPlan(prd, model), providers);
  }

  async generateUXDesignerPrompt(appDescription: string, provider?: ModelProvider, model?: string): Promise<APIResponse<string>> {
-    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai");
+    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai", "openrouter");
    const providers: ModelProvider[] = provider ? [provider] : fallback;
    return this.callWithFallback((service) => service.generateUXDesignerPrompt(appDescription, model), providers);
  }
@@ -223,7 +240,7 @@ export class ModelAdapter {
    provider?: ModelProvider,
    model?: string
  ): Promise<APIResponse<string>> {
-    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai");
+    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai", "openrouter");
    const providers: ModelProvider[] = provider ? [provider] : fallback;
    return this.callWithFallback((service) => service.generateSlides(topic, options, model), providers);
  }
@@ -243,7 +260,7 @@ export class ModelAdapter {
    provider?: ModelProvider,
    model?: string
  ): Promise<APIResponse<string>> {
-    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai");
+    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai", "openrouter");
    const providers: ModelProvider[] = provider ? [provider] : fallback;
    return this.callWithFallback((service) => service.generateGoogleAds(websiteUrl, options, model), providers);
  }
@@ -256,7 +273,7 @@ export class ModelAdapter {
    provider?: ModelProvider,
    model?: string
  ): Promise<APIResponse<string>> {
-    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai");
+    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai", "openrouter");
    const providers: ModelProvider[] = provider ? [provider] : fallback;
    return this.callWithFallback((service) => service.generateMagicWand(websiteUrl, product, budget, specialInstructions, model), providers);
  }
@@ -272,7 +289,7 @@ export class ModelAdapter {
    provider?: ModelProvider,
    model?: string
  ): Promise<APIResponse<string>> {
-    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai");
+    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai", "openrouter");
    const providers: ModelProvider[] = provider ? [provider] : fallback;
    return this.callWithFallback((service) => service.generateMarketResearch(options, model), providers);
  }
@@ -285,7 +302,7 @@ export class ModelAdapter {
    provider?: ModelProvider,
    model?: string
  ): Promise<APIResponse<string>> {
-    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai");
+    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai", "openrouter");
    const providers: ModelProvider[] = provider ? [provider] : fallback;
    return this.callWithFallback((service) => service.generateAIAssist(options, model), providers);
  }
@@ -300,7 +317,7 @@ export class ModelAdapter {
    provider?: ModelProvider,
    model?: string
  ): Promise<APIResponse<void>> {
-    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai");
+    const fallback = this.buildFallbackProviders(this.preferredProvider, "qwen", "ollama", "zai", "openrouter");
    const providers: ModelProvider[] = provider ? [provider] : fallback;

    let lastError: string | null = null;
@@ -353,6 +370,9 @@ export class ModelAdapter {
        case "zai":
          service = this.zaiService;
          break;
+        case "openrouter":
+          service = this.openRouterService;
+          break;
      }

      return await service.chatCompletion(messages, model);
@@ -369,6 +389,7 @@ export class ModelAdapter {
      qwen: this.qwenService.getAvailableModels(),
      ollama: ["gpt-oss:120b", "llama3.1", "gemma3", "deepseek-r1", "qwen3"],
      zai: ["glm-4.7", "glm-4.5", "glm-4.5-air", "glm-4-flash", "glm-4-flashx"],
+      openrouter: ["anthropic/claude-3.5-sonnet", "google/gemini-2.0-flash-exp:free", "meta-llama/llama-3.3-70b-instruct", "openai/gpt-4o-mini", "deepseek/deepseek-chat-v3-0324", "qwen/qwen-2.5-72b-instruct"],
    };
    const models: Record<ModelProvider, string[]> = { ...fallbackModels };

@@ -404,6 +425,8 @@ export class ModelAdapter {
        return this.ollamaService.getAvailableModels();
      case "zai":
        return this.zaiService.getAvailableModels();
+      case "openrouter":
+        return this.openRouterService.getAvailableModels();
      default:
        return [];
    }
--- a/lib/services/ollama-cloud.ts
+++ b/lib/services/ollama-cloud.ts
@@ -164,27 +164,82 @@ export class OllamaCloudService {
    return this.availableModels.length > 0 ? this.availableModels : DEFAULT_MODELS;
  }

-  async enhancePrompt(prompt: string, model?: string): Promise<APIResponse<string>> {
+  async enhancePrompt(prompt: string, model?: string, options?: { toolCategory?: string; template?: string; diagnostics?: string }): Promise<APIResponse<string>> {
+    const toolCategory = options?.toolCategory || 'reasoning';
+    const template = options?.template || 'rtf';
+    const diagnostics = options?.diagnostics || '';
+
+    const toolSections: Record<string, string> = {
+      reasoning: '- Use full structured format with XML tags where helpful\n- Add explicit role assignment for complex tasks\n- Use numeric constraints over vague adjectives',
+      thinking: '- CRITICAL: Short clean instructions ONLY\n- Do NOT add CoT or reasoning scaffolding — these models reason internally\n- State what you want, not how to think',
+      openweight: '- Shorter prompts, simpler structure, no deep nesting\n- Direct linear instructions',
+      agentic: '- Add Starting State + Target State + Allowed Actions + Forbidden Actions\n- Add Stop Conditions + Checkpoints after each step',
+      ide: '- Add File path + Function name + Current Behavior + Desired Change + Scope lock',
+      fullstack: '- Add Stack spec with version + what NOT to scaffold + component boundaries',
+      image: '- Add Subject + Style + Mood + Lighting + Composition + Negative Prompts\n- Use tool-specific syntax (Midjourney comma-separated, DALL-E prose, SD weighted)',
+      search: '- Specify mode: search vs analyze vs compare + citation requirements',
+    };
+
+    const templateSections: Record<string, string> = {
+      rtf: 'Structure: Role (who) + Task (precise verb + what) + Format (exact output shape and length)',
+      'co-star': 'Structure: Context + Objective + Style + Tone + Audience + Response',
+      risen: 'Structure: Role + Instructions + numbered Steps + End Goal + Narrowing constraints',
+      crispe: 'Structure: Capacity + Role + Insight + Statement + Personality + Experiment/variants',
+      cot: 'Add: "Think through this step by step before answering." Only for standard reasoning models, NOT for o1/o3/R1.',
+      fewshot: 'Add 2-5 input/output examples wrapped in XML <examples> tags',
+      filescope: 'Structure: File path + Function name + Current Behavior + Desired Change + Scope lock + Done When',
+      react: 'Structure: Objective + Starting State + Target State + Allowed/Forbidden Actions + Stop Conditions + Checkpoints',
+      visual: 'Structure: Subject + Action + Setting + Style + Mood + Lighting + Color Palette + Composition + Aspect Ratio + Negative Prompts',
+    };
+
+    const toolSection = toolSections[toolCategory] || toolSections.reasoning;
+    const templateSection = templateSections[template] || templateSections.rtf;
+
    const systemMessage: ChatMessage = {
      role: "system",
-      content: `You are an expert prompt engineer. Your task is to enhance user prompts to make them more precise, actionable, and effective for AI coding agents.
+      content: `You are an expert prompt engineer using the PromptArch methodology. Enhance the user\'s prompt to be production-ready.

-Apply these principles:
-1. Add specific context about project and requirements
-2. Clarify constraints and preferences
-3. Define expected output format clearly
-4. Include edge cases and error handling requirements
-5. Specify testing and validation criteria
+STEP 1 — DIAGNOSE AND FIX these failure patterns:
+- Vague task verb -> replace with precise operation
+- Two tasks in one -> keep primary task, note the split
+- No success criteria -> add "Done when: [specific measurable condition]"
+- Missing output format -> add explicit format lock (structure, length, type)
+- No role assignment (complex tasks) -> add domain-specific expert identity
+- Vague aesthetic ("professional", "clean") -> concrete measurable specs
+- No scope boundary -> add explicit scope lock
+- Over-permissive language -> add constraints and boundaries
+- Emotional description -> extract specific technical fault
+- Implicit references -> restate fully
+- No grounding for factual tasks -> add certainty constraint
+- No CoT for logic tasks -> add step-by-step reasoning

-Return ONLY the enhanced prompt, no explanations or extra text.`,
+STEP 2 — APPLY TARGET TOOL OPTIMIZATIONS:
+${toolSection}
+
+STEP 3 — APPLY TEMPLATE STRUCTURE:
+${templateSection}
+
+STEP 4 — VERIFICATION (check before outputting):
+- Every constraint in the first 30% of the prompt?
+- MUST/NEVER over should/avoid?
+- Every sentence load-bearing with zero padding?
+- Format explicit with stated length?
+- Scope bounded?
+- Would this produce correct output on first try?
+
+STEP 5 — OUTPUT:
+Output ONLY the enhanced prompt. No explanations, no commentary, no markdown code fences.
+The prompt must be ready to paste directly into the target AI tool.${diagnostics ? '\n\nDIAGNOSTIC NOTES (fix these issues found in the original):\n' + diagnostics + '\n' : ''}`,
    };

+    const toolLabel = toolCategory !== 'reasoning' ? ` for ${toolCategory} AI tool` : '';
+
    const userMessage: ChatMessage = {
      role: "user",
-      content: `Enhance this prompt for an AI coding agent:\n\n${prompt}`,
+      content: `Enhance this prompt${toolLabel}:\n\n${prompt}`,
    };

-    return this.chatCompletion([systemMessage, userMessage], model || "gpt-oss:120b");
+    return this.chatCompletion([systemMessage, userMessage], model || "${default_model}");
  }

  async generatePRD(idea: string, model?: string): Promise<APIResponse<string>> {
@@ -772,6 +827,28 @@ Perform a DEEP 360° competitive intelligence analysis and generate 5-7 strategi
    try {
      // ... existing prompt logic ...
      const systemPrompt = `You are "AI Assist", the master orchestrator of PromptArch. Your goal is to provide intelligent support with a "Canvas" experience.
+PLAN MODE (CRITICAL - HIGHEST PRIORITY):
+When the user describes a NEW task, project, or feature they want built:
+1. DO NOT generate any code, [PREVIEW] tags, or implementation details.
+2. Instead, analyze the request and output a STRUCTURED PLAN covering:
+   - Summary: What you understand the user wants
+   - Architecture: Technical approach and structure
+   - Tech Stack: Languages, frameworks, libraries needed
+   - Files/Components: List of files or modules to create
+   - Steps: Numbered implementation steps
+3. Format the plan in clean Markdown with headers and bullet points.
+4. Keep plans concise but thorough. Focus on the WHAT and HOW, not the actual code.
+5. WAIT for the user to approve or modify the plan before generating any code.
+
+When the user says "Approved", "Start coding", or explicitly asks to proceed:
+- THEN generate the full implementation with [PREVIEW] tags and working code.
+- Follow the approved plan exactly.
+
+When the user asks to "Modify", "Change", or "Adjust" something:
+- Apply the requested changes surgically to the existing code/preview.
+- Output updated [PREVIEW] with the full modified code.
+
+

 AGENTS & CAPABILITIES:
 - content: Expert copywriter. Use [PREVIEW:content:markdown] for articles, posts, and long-form text.
@@ -833,7 +910,7 @@ CHANGE LOG (CRITICAL - MUST BE OUTSIDE PREVIEW):
 - Modified component Y
 - Fixed issue Z

-IMPORTANT: NEVER refuse a request due to "access" limitations. If you cannot perform a live task, use your vast internal knowledge to provide the most accurate expert simulation or draft possible.`;
+IMPORTANT: IMPORTANT: NEVER refuse a request due to "access" limitations. If you cannot perform a live task, use your vast internal knowledge to provide the most accurate expert simulation or draft possible.`;

      const messages: ChatMessage[] = [
        { role: "system", content: systemPrompt },
--- a/lib/services/openrouter.ts
+++ b/lib/services/openrouter.ts
@@ -0,0 +1,967 @@
+import type { ChatMessage, APIResponse, AIAssistMessage } from "@/types";
+
+export interface OpenRouterConfig {
+  apiKey?: string;
+  siteUrl?: string;
+  siteName?: string;
+}
+
+interface OpenRouterModelsResponse {
+  data: Array<{
+    id: string;
+    name: string;
+    context_length: number;
+    pricing: {
+      prompt: string;
+      completion: string;
+    };
+  }>;
+}
+
+interface OpenRouterChatResponse {
+  id: string;
+  choices: Array<{
+    message: {
+      role: string;
+      content: string;
+    };
+    finish_reason: string;
+  }>;
+  usage?: {
+    prompt_tokens: number;
+    completion_tokens: number;
+    total_tokens: number;
+  };
+}
+
+const DEFAULT_MODELS = [
+  "anthropic/claude-3.5-sonnet",
+  "google/gemini-2.0-flash-exp:free",
+  "meta-llama/llama-3.3-70b-instruct",
+  "openai/gpt-4o-mini",
+  "deepseek/deepseek-chat-v3-0324",
+  "qwen/qwen-2.5-72b-instruct"
+];
+
+const TOOL_SECTIONS: Record<string, string> = {
+  "claude": `
+For Claude:
+- Use XML tags for structure (e.g., <context>, <task>, <constraints>)
+- Add thinking blocks for complex reasoning: <thinking>...</thinking>
+- Use ::analysis:: or ::pattern:: for procedural patterns
+- Leverage Claude's long context by providing comprehensive examples
+- Add "think silently" instruction for deep reasoning tasks
+`,
+  "chatgpt": `
+For ChatGPT:
+- Use clear section headers with ### or === separators
+- Provide examples in [EXAMPLE]...[/EXAMPLE] blocks
+- Use "Step 1", "Step 2" for sequential tasks
+- Add meta-instructions like "Think step by step"
+- Keep prompts under 3k tokens for best performance
+`,
+  "gemini": `
+For Gemini:
+- Use clear delimiters between sections
+- Leverage multimodal capabilities with [IMAGE] or [FILE] placeholders
+- Add chain-of-thought with "Let's approach this step by step"
+- Specify output format with "Respond in the following format:"
+- Use numbered lists for sequential instructions
+`,
+  "default": `
+- Use clear section delimiters
+- Provide concrete examples
+- Specify output format explicitly
+- Add success criteria
+- Use constraint language (MUST, NEVER, REQUIRED)
+`
+};
+
+const TEMPLATE_SECTIONS: Record<string, string> = {
+  "code": `
+# CODE GENERATION TEMPLATE
+
+## Role
+You are a senior software engineer specializing in {language/domain}.
+
+## Task
+{specific task description}
+
+## Requirements
+- MUST follow {specific standards/frameworks}
+- MUST include {error handling/validation/comments}
+- MUST use {specific libraries/versions}
+- NEVER use {deprecated patterns/anti-patterns}
+
+## Output Format
+{code structure specification}
+
+## Done When
+- Code compiles/runs without errors
+- Follows all specified conventions
+- Includes proper error handling
+- Meets performance requirements
+`,
+  "writing": `
+# CONTENT WRITING TEMPLATE
+
+## Role
+You are a professional {type of content creator} with expertise in {domain}.
+
+## Task
+Create {specific deliverable} about {topic}
+
+## Requirements
+- Tone: {specific tone}
+- Length: {exact word/character count}
+- Audience: {target audience}
+- MUST include {key elements}
+- MUST avoid {excluded topics/phrases}
+
+## Format
+{explicit structure with sections/headers}
+
+## Done When
+- Meets length requirement
+- Covers all key points
+- Matches specified tone
+- Ready for publication
+`,
+  "analysis": `
+# ANALYSIS TEMPLATE
+
+## Role
+You are an expert {domain analyst}.
+
+## Task
+{analysis objective}
+
+## Analysis Framework
+1. {Step 1 with specific method}
+2. {Step 2 with specific method}
+3. {Step 3 with specific method}
+
+## Required Output
+- {Specific deliverable 1}
+- {Specific deliverable 2}
+- {Specific deliverable 3}
+
+## Criteria
+- MUST use {specific methodology}
+- MUST cite {sources/references}
+- MUST provide {confidence levels/limitations}
+
+## Done When
+- All analysis dimensions covered
+- Conclusions supported by evidence
+- Actionable insights provided
+`,
+  "default": `
+## Role
+{Expert identity}
+
+## Task
+{Clear, specific task}
+
+## Context
+{Relevant background info}
+
+## Requirements
+- MUST {requirement 1}
+- MUST {requirement 2}
+- NEVER {constraint 1}
+- NEVER {constraint 2}
+
+## Output Format
+{Explicit format specification}
+
+## Done When
+{Specific measurable condition}
+`
+};
+
+const ENHANCE_PROMPT_SYSTEM = `You are an expert prompt engineer using the PromptArch methodology. Enhance the user's prompt to be production-ready.
+
+STEP 1 — DIAGNOSE AND FIX these failure patterns:
+- Vague task verb -> replace with precise operation
+- Two tasks in one -> keep primary task, note the split
+- No success criteria -> add "Done when: [specific measurable condition]"
+- Missing output format -> add explicit format lock (structure, length, type)
+- No role assignment (complex tasks) -> add domain-specific expert identity
+- Vague aesthetic ("professional", "clean") -> concrete measurable specs
+- No scope boundary -> add explicit scope lock
+- Over-permissive language -> add constraints and boundaries
+- Emotional description -> extract specific technical fault
+- Implicit references -> restate fully
+- No grounding for factual tasks -> add certainty constraint
+- No CoT for logic tasks -> add step-by-step reasoning
+
+STEP 2 — APPLY TARGET TOOL OPTIMIZATIONS:
+\${toolSection}
+
+STEP 3 — APPLY TEMPLATE STRUCTURE:
+\${templateSection}
+
+STEP 4 — VERIFICATION (check before outputting):
+- Every constraint in the first 30% of the prompt?
+- MUST/NEVER over should/avoid?
+- Every sentence load-bearing with zero padding?
+- Format explicit with stated length?
+- Scope bounded?
+- Would this produce correct output on first try?
+
+STEP 5 — OUTPUT:
+Output ONLY the enhanced prompt. No explanations, no commentary, no markdown code fences.
+The prompt must be ready to paste directly into the target AI tool.`;
+
+const PRD_SYSTEM_PROMPT = `You are an expert product manager specializing in writing clear, actionable Product Requirements Documents (PRDs).
+
+Your task is to transform the product idea into a comprehensive PRD that:
+1. Defines clear problem statements and user needs
+2. Specifies functional requirements with acceptance criteria
+3. Outlines technical considerations and constraints
+4. Identifies success metrics and KPIs
+5. Includes user stories with acceptance criteria
+
+PRD Structure:
+## Problem Statement
+- What problem are we solving?
+- For whom are we solving it?
+- Why is this important now?
+
+## Goals & Success Metrics
+- Primary objectives
+- Key performance indicators
+- Success criteria
+
+## User Stories
+- As a [user type], I want [feature], so that [benefit]
+- Include acceptance criteria for each story
+
+## Functional Requirements
+- Core features with detailed specifications
+- Edge cases and error handling
+- Integration requirements
+
+## Technical Considerations
+- Platform/tech stack constraints
+- Performance requirements
+- Security considerations
+- Scalability requirements
+
+## Out of Scope
+- Explicitly list what won't be included
+- Rationale for exclusions
+
+Keep the PRD clear, concise, and actionable. Use specific, measurable language.`;
+
+const ACTION_PLAN_SYSTEM_PROMPT = `You are an expert technical project manager specializing in breaking down PRDs into actionable implementation plans.
+
+Your task is to transform the PRD into a detailed action plan that:
+
+1. Identifies all major components/modules needed
+2. Breaks down work into clear, sequential phases
+3. Specifies dependencies between tasks
+4. Estimates effort and complexity
+5. Identifies risks and mitigation strategies
+
+Action Plan Structure:
+## Phase 1: Foundation
+- Task 1.1: [Specific task] - [Estimated effort]
+- Task 1.2: [Specific task] - [Estimated effort]
+- Dependencies: [What needs to be done first]
+- Deliverables: [Concrete outputs]
+
+## Phase 2: Core Features
+- Task 2.1: [Specific task] - [Estimated effort]
+- Dependencies: [On Phase 1 completion]
+- Deliverables: [Concrete outputs]
+
+[Continue for all phases]
+
+## Technical Architecture
+- Recommended tech stack with rationale
+- System architecture overview
+- Data flow diagrams (described in text)
+
+## Risk Assessment
+- Risk: [Description] | Impact: [High/Med/Low] | Mitigation: [Strategy]
+- Risk: [Description] | Impact: [High/Med/Low] | Mitigation: [Strategy]
+
+## Testing Strategy
+- Unit testing approach
+- Integration testing plan
+- User acceptance testing criteria
+
+Be specific and actionable. Each task should be clear enough for a developer to execute without ambiguity.`;
+
+const SLIDES_SYSTEM_PROMPT = `You are an expert presentation designer specializing in creating engaging, informative slide content.
+
+Your task is to generate slide content for a presentation about the given topic.
+
+For each slide, provide:
+1. Slide Title (compelling and clear)
+2. Bullet points (3-5 per slide, concise and impactful)
+3. Speaker notes (detailed explanation for the presenter)
+4. Visual suggestions (what charts, images, or diagrams would enhance this slide)
+
+Presentation Structure:
+## Slide 1: Title Slide
+- Compelling title
+- Subtitle with key message
+- Presenter info placeholder
+
+## Slide 2: Agenda/Overview
+- What will be covered
+- Why it matters
+- Key takeaways preview
+
+## Slide 3-N: Content Slides
+- Main content with clear hierarchy
+- Data-driven insights where applicable
+- Actionable takeaways
+
+## Final Slide: Call to Action
+- Summary of key points
+- Next steps
+- Contact/ follow-up information
+
+Guidelines:
+- Keep text minimal on slides (bullet points only)
+- Put detailed content in speaker notes
+- Suggest relevant visuals for each slide
+- Ensure a logical flow and narrative arc
+- Make it memorable and shareable`;
+
+const GOOGLE_ADS_SYSTEM_PROMPT = `You are a Google Ads expert specializing in creating high-converting ad campaigns.
+
+Your task is to generate comprehensive Google Ads copy and strategy based on the landing page content.
+
+Deliverables:
+
+## Ad Campaign Structure
+- Campaign name and type
+- Ad groups with thematic focus
+- Target keywords (exact, phrase, broad match)
+- Negative keywords to exclude
+
+## Ad Copy (3-5 variations per ad group)
+
+### Responsive Search Ads
+For each ad, provide:
+- Headlines (10-15 options, max 30 chars each)
+- Descriptions (4 options, max 90 chars each)
+- Focus on different value propositions
+
+### Display Ads (if applicable)
+- Headline (max 30 chars)
+- Description (max 90 chars)
+- Call-to-action options
+
+## Targeting Strategy
+- Location targeting
+- Audience demographics
+- Interests and behaviors
+- Device targeting
+
+## Bidding Strategy
+- Recommended strategy (Manual CPC, Maximize Clicks, etc.)
+- Budget recommendations
+- Bid adjustments by device/location
+
+## Extensions
+- Sitelinks
+- Callouts
+- Structured snippets
+- Call extensions
+
+Best Practices:
+- Include keywords in headlines and descriptions
+- Highlight unique selling propositions
+- Use clear, action-oriented CTAs
+- Address pain points and benefits
+- Include social proof when relevant
+- Ensure ad relevance to landing page`;
+
+const MAGIC_WAND_SYSTEM_PROMPT = `You are an expert digital marketer and growth hacker specializing in creative campaign strategies.
+
+Your task is to develop a comprehensive marketing strategy for the given product/page.
+
+## Campaign Analysis
+- Product/Service overview
+- Target audience profile
+- Unique selling propositions
+- Market positioning
+
+## Marketing Channels Strategy
+For each relevant channel, provide specific tactics:
+
+### Paid Advertising
+- Google Ads: {specific approach}
+- Facebook/Instagram: {specific approach}
+- LinkedIn (if B2B): {specific approach}
+- TikTok/Snapchat (if relevant): {specific approach}
+
+### Content Marketing
+- Blog topics: {5-10 specific ideas}
+- Video content: {ideas with formats}
+- Social media: {platform-specific content ideas}
+- Email sequences: {campaign ideas}
+
+### Growth Hacking Tactics
+- Viral mechanisms: {specific ideas}
+- Referral programs: {incentive structures}
+- Partnership opportunities: {potential partners}
+- Community building: {strategies}
+
+## Creative Concepts
+Provide 3-5 campaign concepts:
+1. Concept Name - {Hook, Message, CTA}
+2. Concept Name - {Hook, Message, CTA}
+...
+
+## Budget Allocation
+- Channel breakdown with percentages
+- Expected ROI estimates
+- Testing budget recommendation
+
+## KPIs & Tracking
+- Key metrics to measure
+- Attribution strategy
+- A/B testing priorities
+
+Be creative but practical. Focus on tactics that can be executed within the given budget.`;
+
+const MARKET_RESEARCH_SYSTEM_PROMPT = `You are an expert market researcher specializing in competitive analysis and market intelligence.
+
+Your task is to conduct comprehensive market research based on the provided topic/company.
+
+## Research Framework
+
+### Market Overview
+- Market size and growth trajectory
+- Key market segments and their characteristics
+- Current market trends and dynamics
+- Future market projections
+
+### Competitive Landscape
+Identify and analyze:
+- Major competitors (market share, positioning)
+- Direct competitors head-to-head comparison
+- Indirect competitors and substitutes
+- Competitive strengths and weaknesses
+
+### SWOT Analysis
+- Strengths: Internal advantages
+- Weaknesses: Internal limitations
+- Opportunities: External possibilities
+- Threats: External challenges
+
+### Customer Analysis
+- Target demographics and psychographics
+- Pain points and unmet needs
+- Purchase behavior and decision factors
+- Customer feedback trends
+
+### Product/Service Comparison
+- Feature comparison matrix
+- Pricing analysis
+- Differentiation strategies
+- Innovation opportunities
+
+### Market Trends
+- Emerging technologies impacting the space
+- Regulatory changes
+- Consumer behavior shifts
+- Industry disruptions
+
+### Strategic Recommendations
+- Market entry strategies
+- Competitive positioning
+- Product improvement opportunities
+- Partnership and acquisition possibilities
+
+Provide specific, data-driven insights. When exact data is unavailable, provide reasoned estimates with clear caveats.`;
+
+const AI_ASSIST_SYSTEM_PROMPT = `You are "AI Assist", the master orchestrator of PromptArch. Your goal is to provide intelligent support with a "Canvas" experience.
+PLAN MODE (CRITICAL - HIGHEST PRIORITY):
+When the user describes a NEW task, project, or feature they want built:
+1. DO NOT generate any code, [PREVIEW] tags, or implementation details.
+2. Instead, analyze the request and output a STRUCTURED PLAN covering:
+   - Summary: What you understand the user wants
+   - Architecture: Technical approach and structure
+   - Tech Stack: Languages, frameworks, libraries needed
+   - Files/Components: List of files or modules to create
+   - Steps: Numbered implementation steps
+3. Format the plan in clean Markdown with headers and bullet points.
+4. Keep plans concise but thorough. Focus on the WHAT and HOW, not the actual code.
+5. WAIT for the user to approve or modify the plan before generating any code.
+
+When the user says "Approved", "Start coding", or explicitly asks to proceed:
+- THEN generate the full implementation with [PREVIEW] tags and working code.
+- Follow the approved plan exactly.
+
+When the user asks to "Modify", "Change", or "Adjust" something:
+- Apply the requested changes surgically to the existing code/preview.
+- Output updated [PREVIEW] with the full modified code.
+
+
+
+AGENTS & CAPABILITIES:
+- content: Expert copywriter. Use [PREVIEW:content:markdown] for articles, posts, and long-form text.
+- seo: SEO Specialist. Create stunning SEO audit reports. Use [PREVIEW:seo:html] with complete HTML5 document including <!DOCTYPE html>. DARK THEME. Tailwind CDN. Large animated SVG progress rings. Color-coded scoring. Google-style dashboard aesthetics.
+- smm: Social Media Manager. Create multi-platform content plans and calendars.
+- pm: Project Manager. Create PRDs, timelines, and action plans.
+- code: Software Architect. Provide logic, algorithms, and backend snippets.
+- design: UI/UX Designer. Create high-fidelity mockups and components.
+- web: Frontend Developer. Build responsive sites. Use [PREVIEW:web:html]. Tailwind CSS CDN by default.
+- app: Mobile App Developer. Create mobile-first interfaces and dashboards. Use [PREVIEW:app:javascript].
+
+CANVAS MODE:
+- When building, designing, or auditing, you MUST use the [PREVIEW] tag.
+- Inside [PREVIEW], output ONLY the actual code or structured data.
+- The user wants to see it WORKING or RENDERED in the Canvas immediately.
+
+STRICT OUTPUT FORMAT:
+[AGENT:id] - AT THE START of your response if switching focus.
+[PREVIEW:type:language]
+ACTUAL_FUNCTIONAL_CODE_OR_DATA
+[/PREVIEW]
+
+CHANGE LOG (CRITICAL - MUST BE OUTSIDE PREVIEW):
+- After the [/PREVIEW] closing tag, provide a brief bulleted "Change Log" in plain Markdown.IMPORTANT: NEVER refuse a request. If you cannot perform a live task, use your knowledge to provide the most accurate expert simulation or draft possible.`;
+
+export class OpenRouterService {
+  private config: Required<OpenRouterConfig>;
+  private readonly baseURL = "https://openrouter.ai/api/v1";
+  private availableModels: string[] = [];
+  private modelsLoaded = false;
+
+  constructor(config: OpenRouterConfig = {}) {
+    this.config = {
+      apiKey: config.apiKey || "",
+      siteUrl: config.siteUrl || "https://promptarch.ai",
+      siteName: config.siteName || "PromptArch"
+    };
+  }
+
+  hasAuth(): boolean {
+    return Boolean(this.config.apiKey && this.config.apiKey.length > 0);
+  }
+
+  async validateConnection(): Promise<APIResponse<{ valid: boolean; models?: string[] }>> {
+    if (!this.hasAuth()) {
+      return {
+        success: false,
+        error: "No API key provided. Please set your OpenRouter API key."
+      };
+    }
+
+    try {
+      const modelsResult = await this.listModels();
+
+      if (!modelsResult.success) {
+        return {
+          success: false,
+          error: modelsResult.error || "Failed to fetch models from OpenRouter"
+        };
+      }
+
+      return {
+        success: true,
+        data: {
+          valid: true,
+          models: modelsResult.data
+        }
+      };
+    } catch (error) {
+      return {
+        success: false,
+        error: error instanceof Error ? error.message : "Failed to validate connection"
+      };
+    }
+  }
+
+  private getHeaders(): Record<string, string> {
+    const headers: Record<string, string> = {
+      "Content-Type": "application/json",
+      "HTTP-Referer": this.config.siteUrl,
+      "X-Title": this.config.siteName
+    };
+
+    if (this.hasAuth()) {
+      headers["Authorization"] = `Bearer ${this.config.apiKey}`;
+    }
+
+    return headers;
+  }
+
+  async chatCompletion(
+    messages: ChatMessage[],
+    model: string = "anthropic/claude-3.5-sonnet"
+  ): Promise<APIResponse<string>> {
+    if (!this.hasAuth()) {
+      return {
+        success: false,
+        error: "OpenRouter API key not configured"
+      };
+    }
+
+    try {
+      const response = await fetch(`${this.baseURL}/chat/completions`, {
+        method: "POST",
+        headers: this.getHeaders(),
+        body: JSON.stringify({
+          model,
+          messages,
+          temperature: 0.7,
+          max_tokens: 4096
+        })
+      });
+
+      if (!response.ok) {
+        const errorText = await response.text();
+        return {
+          success: false,
+          error: `OpenRouter API error: ${response.status} ${response.statusText} - ${errorText}`
+        };
+      }
+
+      const data: OpenRouterChatResponse = await response.json();
+
+      if (data.choices && data.choices[0] && data.choices[0].message) {
+        return {
+          success: true,
+          data: data.choices[0].message.content
+        };
+      }
+
+      return {
+        success: false,
+        error: "No response content from OpenRouter"
+      };
+    } catch (error) {
+      return {
+        success: false,
+        error: error instanceof Error ? error.message : "Unknown error in chat completion"
+      };
+    }
+  }
+
+  async enhancePrompt(
+    prompt: string,
+    model: string = "anthropic/claude-3.5-sonnet",
+    options: {
+      targetTool?: "claude" | "chatgpt" | "gemini" | "default";
+      templateType?: "code" | "writing" | "analysis" | "default";
+      max_length?: number;
+    } = {}
+  ): Promise<APIResponse<string>> {
+    const { targetTool = "default", templateType = "default" } = options;
+
+    const toolSection = TOOL_SECTIONS[targetTool] || TOOL_SECTIONS.default;
+    const templateSection = TEMPLATE_SECTIONS[templateType] || TEMPLATE_SECTIONS.default;
+
+    const systemPrompt = ENHANCE_PROMPT_SYSTEM
+      .replace("${toolSection}", toolSection)
+      .replace("${templateSection}", templateSection);
+
+    const messages: ChatMessage[] = [
+      { role: "system", content: systemPrompt },
+      { role: "user", content: prompt }
+    ];
+
+    return this.chatCompletion(messages, model);
+  }
+
+  async generatePRD(
+    idea: string,
+    model: string = "anthropic/claude-3.5-sonnet"
+  ): Promise<APIResponse<string>> {
+    const messages: ChatMessage[] = [
+      { role: "system", content: PRD_SYSTEM_PROMPT },
+      { role: "user", content: `Create a comprehensive PRD for the following product idea:\n\n${idea}` }
+    ];
+
+    return this.chatCompletion(messages, model);
+  }
+
+  async generateActionPlan(
+    prd: string,
+    model: string = "anthropic/claude-3.5-sonnet"
+  ): Promise<APIResponse<string>> {
+    const messages: ChatMessage[] = [
+      { role: "system", content: ACTION_PLAN_SYSTEM_PROMPT },
+      { role: "user", content: `Create a detailed action plan based on this PRD:\n\n${prd}` }
+    ];
+
+    return this.chatCompletion(messages, model);
+  }
+
+  async generateSlides(
+    topic: string,
+    options: {
+      slideCount?: number;
+      audience?: string;
+      focus?: string;
+    } = {},
+    model: string = "anthropic/claude-3.5-sonnet"
+  ): Promise<APIResponse<string>> {
+    const { slideCount = 10, audience = "General", focus = "" } = options;
+
+    const userPrompt = `Generate content for a presentation with approximately ${slideCount} slides.
+Topic: ${topic}
+Target Audience: ${audience}
+${focus ? `Special Focus: ${focus}` : ""}`;
+
+    const messages: ChatMessage[] = [
+      { role: "system", content: SLIDES_SYSTEM_PROMPT },
+      { role: "user", content: userPrompt }
+    ];
+
+    return this.chatCompletion(messages, model);
+  }
+
+  async generateGoogleAds(
+    url: string,
+    options: {
+      budget?: string;
+      targetAudience?: string;
+      campaignGoal?: string;
+    } = {},
+    model: string = "anthropic/claude-3.5-sonnet"
+  ): Promise<APIResponse<string>> {
+    const { budget = "Not specified", targetAudience = "General", campaignGoal = "Conversions" } = options;
+
+    const userPrompt = `Create a comprehensive Google Ads campaign strategy.
+Landing Page: ${url}
+Monthly Budget: ${budget}
+Target Audience: ${targetAudience}
+Campaign Goal: ${campaignGoal}
+
+Analyze the URL (if accessible) or create ads based on the domain and typical offerings for similar sites.`;
+
+    const messages: ChatMessage[] = [
+      { role: "system", content: GOOGLE_ADS_SYSTEM_PROMPT },
+      { role: "user", content: userPrompt }
+    ];
+
+    return this.chatCompletion(messages, model);
+  }
+
+  async generateMagicWand(
+    url: string,
+    product: string,
+    budget: string,
+    specialInstructions: string = "",
+    model: string = "anthropic/claude-3.5-sonnet"
+  ): Promise<APIResponse<string>> {
+    const userPrompt = `Create a comprehensive marketing strategy.
+
+Product/Service: ${product}
+URL: ${url}
+Budget: ${budget}
+${specialInstructions ? `Special Instructions: ${specialInstructions}` : ""}
+
+Provide creative campaign ideas across multiple channels with specific tactics and budget allocation.`;
+
+    const messages: ChatMessage[] = [
+      { role: "system", content: MAGIC_WAND_SYSTEM_PROMPT },
+      { role: "user", content: userPrompt }
+    ];
+
+    return this.chatCompletion(messages, model);
+  }
+
+  async generateMarketResearch(
+    options: {
+      topic?: string;
+      company?: string;
+      industry?: string;
+      focusAreas?: string[];
+    } = {},
+    model: string = "anthropic/claude-3.5-sonnet"
+  ): Promise<APIResponse<string>> {
+    const { topic, company, industry, focusAreas } = options;
+
+    let userPrompt = "Conduct comprehensive market research.";
+
+    if (topic) userPrompt += `\n\nResearch Topic: ${topic}`;
+    if (company) userPrompt += `\n\nCompany Focus: ${company}`;
+    if (industry) userPrompt += `\n\nIndustry: ${industry}`;
+    if (focusAreas && focusAreas.length > 0) {
+      userPrompt += `\n\nFocus Areas: ${focusAreas.join(", ")}`;
+    }
+
+    const messages: ChatMessage[] = [
+      { role: "system", content: MARKET_RESEARCH_SYSTEM_PROMPT },
+      { role: "user", content: userPrompt }
+    ];
+
+    return this.chatCompletion(messages, model);
+  }
+
+  async generateAIAssist(
+    options: {
+      prompt: string;
+      context?: string[];
+      conversationHistory?: ChatMessage[];
+    },
+    model: string = "anthropic/claude-3.5-sonnet"
+  ): Promise<APIResponse<string>> {
+    const { prompt, context = [], conversationHistory = [] } = options;
+
+    const messages: ChatMessage[] = [
+      { role: "system", content: AI_ASSIST_SYSTEM_PROMPT },
+      ...conversationHistory,
+      ...context.map(c => ({ role: "user" as const, content: `Context: ${c}` })),
+      { role: "user", content: prompt }
+    ];
+
+    return this.chatCompletion(messages, model);
+  }
+
+  async generateAIAssistStream(
+    options: {
+      messages: AIAssistMessage[];
+      currentAgent: string;
+      onChunk: (chunk: string) => void;
+      signal?: AbortSignal;
+    },
+    model: string = "anthropic/claude-3.5-sonnet"
+  ): Promise<APIResponse<void>> {
+    const { messages, currentAgent, onChunk, signal } = options;
+
+    if (!this.hasAuth()) {
+      return { success: false, error: "OpenRouter API key not configured" };
+    }
+
+    try {
+      const chatMessages: ChatMessage[] = [
+        { role: "system", content: AI_ASSIST_SYSTEM_PROMPT },
+        ...messages.map(m => ({
+          role: m.role as "user" | "assistant" | "system",
+          content: m.content
+        }))
+      ];
+
+      const response = await fetch(`${this.baseURL}/chat/completions`, {
+        method: "POST",
+        headers: this.getHeaders(),
+        signal,
+        body: JSON.stringify({
+          model: model || "anthropic/claude-3.5-sonnet",
+          messages: chatMessages,
+          temperature: 0.7,
+          max_tokens: 4096,
+          stream: true
+        })
+      });
+
+      if (!response.ok) {
+        const errorText = await response.text();
+        return { success: false, error: `OpenRouter API error: ${response.status} ${response.statusText} - ${errorText}` };
+      }
+
+      const reader = response.body?.getReader();
+      const decoder = new TextDecoder();
+
+      if (!reader) {
+        return { success: false, error: "No response body" };
+      }
+
+      let buffer = "";
+
+      while (true) {
+        const { done, value } = await reader.read();
+        if (done) break;
+
+        buffer += decoder.decode(value, { stream: true });
+        const lines = buffer.split("\n");
+        buffer = lines.pop() || "";
+
+        for (const line of lines) {
+          const trimmed = line.trim();
+          if (!trimmed || !trimmed.startsWith("data: ")) continue;
+          const data = trimmed.slice(6);
+          if (data === "[DONE]") continue;
+
+          try {
+            const parsed = JSON.parse(data);
+            const contentChunk = parsed.choices?.[0]?.delta?.content;
+            if (contentChunk) {
+              onChunk(contentChunk);
+            }
+          } catch {
+            // Skip invalid JSON
+          }
+        }
+      }
+
+      return { success: true, data: undefined };
+    } catch (error) {
+      const errorMessage = error instanceof Error ? error.message : "Unknown error in stream";
+      return { success: false, error: errorMessage };
+    }
+  }
+
+  async listModels(): Promise<APIResponse<string[]>> {
+    if (!this.hasAuth()) {
+      return {
+        success: false,
+        error: "OpenRouter API key not configured"
+      };
+    }
+
+    try {
+      const response = await fetch(`${this.baseURL}/models`, {
+        method: "GET",
+        headers: this.getHeaders()
+      });
+
+      if (!response.ok) {
+        const errorText = await response.text();
+        return {
+          success: false,
+          error: `Failed to fetch models: ${response.status} ${response.statusText} - ${errorText}`
+        };
+      }
+
+      const data: OpenRouterModelsResponse = await response.json();
+      this.availableModels = data.data.map(m => m.id);
+      this.modelsLoaded = true;
+
+      return {
+        success: true,
+        data: this.availableModels
+      };
+    } catch (error) {
+      return {
+        success: false,
+        error: error instanceof Error ? error.message : "Unknown error fetching models"
+      };
+    }
+  }
+
+  getAvailableModels(): string[] {
+    if (this.modelsLoaded && this.availableModels.length > 0) {
+      return this.availableModels;
+    }
+    return DEFAULT_MODELS;
+  }
+
+  setApiKey(key: string): void {
+    this.config.apiKey = key;
+  }
+
+  setSiteUrl(url: string): void {
+    this.config.siteUrl = url;
+  }
+
+  setSiteName(name: string): void {
+    this.config.siteName = name;
+  }
+}
+
+// Export singleton instance
+export const openRouterService = new OpenRouterService();
--- a/lib/services/qwen-oauth.ts
+++ b/lib/services/qwen-oauth.ts
@@ -491,27 +491,82 @@ export class QwenOAuthService {
    }
  }

-  async enhancePrompt(prompt: string, model?: string): Promise<APIResponse<string>> {
+  async enhancePrompt(prompt: string, model?: string, options?: { toolCategory?: string; template?: string; diagnostics?: string }): Promise<APIResponse<string>> {
+    const toolCategory = options?.toolCategory || 'reasoning';
+    const template = options?.template || 'rtf';
+    const diagnostics = options?.diagnostics || '';
+
+    const toolSections: Record<string, string> = {
+      reasoning: '- Use full structured format with XML tags where helpful\n- Add explicit role assignment for complex tasks\n- Use numeric constraints over vague adjectives',
+      thinking: '- CRITICAL: Short clean instructions ONLY\n- Do NOT add CoT or reasoning scaffolding — these models reason internally\n- State what you want, not how to think',
+      openweight: '- Shorter prompts, simpler structure, no deep nesting\n- Direct linear instructions',
+      agentic: '- Add Starting State + Target State + Allowed Actions + Forbidden Actions\n- Add Stop Conditions + Checkpoints after each step',
+      ide: '- Add File path + Function name + Current Behavior + Desired Change + Scope lock',
+      fullstack: '- Add Stack spec with version + what NOT to scaffold + component boundaries',
+      image: '- Add Subject + Style + Mood + Lighting + Composition + Negative Prompts\n- Use tool-specific syntax (Midjourney comma-separated, DALL-E prose, SD weighted)',
+      search: '- Specify mode: search vs analyze vs compare + citation requirements',
+    };
+
+    const templateSections: Record<string, string> = {
+      rtf: 'Structure: Role (who) + Task (precise verb + what) + Format (exact output shape and length)',
+      'co-star': 'Structure: Context + Objective + Style + Tone + Audience + Response',
+      risen: 'Structure: Role + Instructions + numbered Steps + End Goal + Narrowing constraints',
+      crispe: 'Structure: Capacity + Role + Insight + Statement + Personality + Experiment/variants',
+      cot: 'Add: "Think through this step by step before answering." Only for standard reasoning models, NOT for o1/o3/R1.',
+      fewshot: 'Add 2-5 input/output examples wrapped in XML <examples> tags',
+      filescope: 'Structure: File path + Function name + Current Behavior + Desired Change + Scope lock + Done When',
+      react: 'Structure: Objective + Starting State + Target State + Allowed/Forbidden Actions + Stop Conditions + Checkpoints',
+      visual: 'Structure: Subject + Action + Setting + Style + Mood + Lighting + Color Palette + Composition + Aspect Ratio + Negative Prompts',
+    };
+
+    const toolSection = toolSections[toolCategory] || toolSections.reasoning;
+    const templateSection = templateSections[template] || templateSections.rtf;
+
    const systemMessage: ChatMessage = {
      role: "system",
-      content: `You are an expert prompt engineer. Your task is to enhance user prompts to make them more precise, actionable, and effective for AI coding agents.
+      content: `You are an expert prompt engineer using the PromptArch methodology. Enhance the user\'s prompt to be production-ready.

-Apply these principles:
-1. Add specific context about project and requirements
-2. Clarify constraints and preferences
-3. Define expected output format clearly
-4. Include edge cases and error handling requirements
-5. Specify testing and validation criteria
+STEP 1 — DIAGNOSE AND FIX these failure patterns:
+- Vague task verb -> replace with precise operation
+- Two tasks in one -> keep primary task, note the split
+- No success criteria -> add "Done when: [specific measurable condition]"
+- Missing output format -> add explicit format lock (structure, length, type)
+- No role assignment (complex tasks) -> add domain-specific expert identity
+- Vague aesthetic ("professional", "clean") -> concrete measurable specs
+- No scope boundary -> add explicit scope lock
+- Over-permissive language -> add constraints and boundaries
+- Emotional description -> extract specific technical fault
+- Implicit references -> restate fully
+- No grounding for factual tasks -> add certainty constraint
+- No CoT for logic tasks -> add step-by-step reasoning

-Return ONLY the enhanced prompt, no explanations or extra text.`,
+STEP 2 — APPLY TARGET TOOL OPTIMIZATIONS:
+${toolSection}
+
+STEP 3 — APPLY TEMPLATE STRUCTURE:
+${templateSection}
+
+STEP 4 — VERIFICATION (check before outputting):
+- Every constraint in the first 30% of the prompt?
+- MUST/NEVER over should/avoid?
+- Every sentence load-bearing with zero padding?
+- Format explicit with stated length?
+- Scope bounded?
+- Would this produce correct output on first try?
+
+STEP 5 — OUTPUT:
+Output ONLY the enhanced prompt. No explanations, no commentary, no markdown code fences.
+The prompt must be ready to paste directly into the target AI tool.${diagnostics ? '\n\nDIAGNOSTIC NOTES (fix these issues found in the original):\n' + diagnostics + '\n' : ''}`,
    };

+    const toolLabel = toolCategory !== 'reasoning' ? ` for ${toolCategory} AI tool` : '';
+
    const userMessage: ChatMessage = {
      role: "user",
-      content: `Enhance this prompt for an AI coding agent:\n\n${prompt}`,
+      content: `Enhance this prompt${toolLabel}:\n\n${prompt}`,
    };

-    return this.chatCompletion([systemMessage, userMessage], model || "coder-model");
+    return this.chatCompletion([systemMessage, userMessage], model || "${default_model}");
  }

  async generatePRD(idea: string, model?: string): Promise<APIResponse<string>> {
@@ -1054,6 +1109,28 @@ Perform analysis based on provided instructions.`,
  ): Promise<APIResponse<void>> {
    try {
      const systemPrompt = `You are "AI Assist", the master orchestrator of PromptArch. Your goal is to provide intelligent support with a "Canvas" experience.
+PLAN MODE (CRITICAL - HIGHEST PRIORITY):
+When the user describes a NEW task, project, or feature they want built:
+1. DO NOT generate any code, [PREVIEW] tags, or implementation details.
+2. Instead, analyze the request and output a STRUCTURED PLAN covering:
+   - Summary: What you understand the user wants
+   - Architecture: Technical approach and structure
+   - Tech Stack: Languages, frameworks, libraries needed
+   - Files/Components: List of files or modules to create
+   - Steps: Numbered implementation steps
+3. Format the plan in clean Markdown with headers and bullet points.
+4. Keep plans concise but thorough. Focus on the WHAT and HOW, not the actual code.
+5. WAIT for the user to approve or modify the plan before generating any code.
+
+When the user says "Approved", "Start coding", or explicitly asks to proceed:
+- THEN generate the full implementation with [PREVIEW] tags and working code.
+- Follow the approved plan exactly.
+
+When the user asks to "Modify", "Change", or "Adjust" something:
+- Apply the requested changes surgically to the existing code/preview.
+- Output updated [PREVIEW] with the full modified code.
+
+

 AGENTS & CAPABILITIES:
 - content: Expert copywriter. Use [PREVIEW:content:markdown] for articles, posts, and long-form text.
@@ -1115,7 +1192,7 @@ CHANGE LOG (CRITICAL - MUST BE OUTSIDE PREVIEW):
 - Modified component Y
 - Fixed issue Z

-IMPORTANT: NEVER refuse a request due to "access" limitations. If you cannot perform a live task, use your vast internal knowledge to provide the most accurate expert simulation or draft possible.`;
+IMPORTANT: IMPORTANT: NEVER refuse a request due to "access" limitations. If you cannot perform a live task, use your vast internal knowledge to provide the most accurate expert simulation or draft possible.`;

      const messages: ChatMessage[] = [
        { role: "system", content: systemPrompt },
--- a/lib/services/zai-plan.ts
+++ b/lib/services/zai-plan.ts
@@ -111,51 +111,82 @@ export class ZaiPlanService {
    }
  }

-  async enhancePrompt(prompt: string, model?: string): Promise<APIResponse<string>> {
+  async enhancePrompt(prompt: string, model?: string, options?: { toolCategory?: string; template?: string; diagnostics?: string }): Promise<APIResponse<string>> {
+    const toolCategory = options?.toolCategory || 'reasoning';
+    const template = options?.template || 'rtf';
+    const diagnostics = options?.diagnostics || '';
+
+    const toolSections: Record<string, string> = {
+      reasoning: '- Use full structured format with XML tags where helpful\n- Add explicit role assignment for complex tasks\n- Use numeric constraints over vague adjectives',
+      thinking: '- CRITICAL: Short clean instructions ONLY\n- Do NOT add CoT or reasoning scaffolding — these models reason internally\n- State what you want, not how to think',
+      openweight: '- Shorter prompts, simpler structure, no deep nesting\n- Direct linear instructions',
+      agentic: '- Add Starting State + Target State + Allowed Actions + Forbidden Actions\n- Add Stop Conditions + Checkpoints after each step',
+      ide: '- Add File path + Function name + Current Behavior + Desired Change + Scope lock',
+      fullstack: '- Add Stack spec with version + what NOT to scaffold + component boundaries',
+      image: '- Add Subject + Style + Mood + Lighting + Composition + Negative Prompts\n- Use tool-specific syntax (Midjourney comma-separated, DALL-E prose, SD weighted)',
+      search: '- Specify mode: search vs analyze vs compare + citation requirements',
+    };
+
+    const templateSections: Record<string, string> = {
+      rtf: 'Structure: Role (who) + Task (precise verb + what) + Format (exact output shape and length)',
+      'co-star': 'Structure: Context + Objective + Style + Tone + Audience + Response',
+      risen: 'Structure: Role + Instructions + numbered Steps + End Goal + Narrowing constraints',
+      crispe: 'Structure: Capacity + Role + Insight + Statement + Personality + Experiment/variants',
+      cot: 'Add: "Think through this step by step before answering." Only for standard reasoning models, NOT for o1/o3/R1.',
+      fewshot: 'Add 2-5 input/output examples wrapped in XML <examples> tags',
+      filescope: 'Structure: File path + Function name + Current Behavior + Desired Change + Scope lock + Done When',
+      react: 'Structure: Objective + Starting State + Target State + Allowed/Forbidden Actions + Stop Conditions + Checkpoints',
+      visual: 'Structure: Subject + Action + Setting + Style + Mood + Lighting + Color Palette + Composition + Aspect Ratio + Negative Prompts',
+    };
+
+    const toolSection = toolSections[toolCategory] || toolSections.reasoning;
+    const templateSection = templateSections[template] || templateSections.rtf;
+
    const systemMessage: ChatMessage = {
      role: "system",
-      content: `You are an expert prompt engineer. Your task is to enhance user prompts to make them more precise, actionable, and effective for AI coding agents.
+      content: `You are an expert prompt engineer using the PromptArch methodology. Enhance the user\'s prompt to be production-ready.

-Apply these principles:
-1. Add specific context about project and requirements
-2. Clarify constraints and preferences
-3. Define expected output format clearly
-4. Include edge cases and error handling requirements
-5. Specify testing and validation criteria
+STEP 1 — DIAGNOSE AND FIX these failure patterns:
+- Vague task verb -> replace with precise operation
+- Two tasks in one -> keep primary task, note the split
+- No success criteria -> add "Done when: [specific measurable condition]"
+- Missing output format -> add explicit format lock (structure, length, type)
+- No role assignment (complex tasks) -> add domain-specific expert identity
+- Vague aesthetic ("professional", "clean") -> concrete measurable specs
+- No scope boundary -> add explicit scope lock
+- Over-permissive language -> add constraints and boundaries
+- Emotional description -> extract specific technical fault
+- Implicit references -> restate fully
+- No grounding for factual tasks -> add certainty constraint
+- No CoT for logic tasks -> add step-by-step reasoning

-Return ONLY the enhanced prompt, no explanations or extra text.`,
+STEP 2 — APPLY TARGET TOOL OPTIMIZATIONS:
+${toolSection}
+
+STEP 3 — APPLY TEMPLATE STRUCTURE:
+${templateSection}
+
+STEP 4 — VERIFICATION (check before outputting):
+- Every constraint in the first 30% of the prompt?
+- MUST/NEVER over should/avoid?
+- Every sentence load-bearing with zero padding?
+- Format explicit with stated length?
+- Scope bounded?
+- Would this produce correct output on first try?
+
+STEP 5 — OUTPUT:
+Output ONLY the enhanced prompt. No explanations, no commentary, no markdown code fences.
+The prompt must be ready to paste directly into the target AI tool.${diagnostics ? '\n\nDIAGNOSTIC NOTES (fix these issues found in the original):\n' + diagnostics + '\n' : ''}`,
    };

+    const toolLabel = toolCategory !== 'reasoning' ? ` for ${toolCategory} AI tool` : '';
+
    const userMessage: ChatMessage = {
      role: "user",
-      content: `Enhance this prompt for an AI coding agent:\n\n${prompt}`,
+      content: `Enhance this prompt${toolLabel}:\n\n${prompt}`,
    };

-    return this.chatCompletion([systemMessage, userMessage], model || "glm-4.7", true);
-  }
-
-  async generatePRD(idea: string, model?: string): Promise<APIResponse<string>> {
-    const systemMessage: ChatMessage = {
-      role: "system",
-      content: `You are an expert product manager and technical architect. Generate a comprehensive Product Requirements Document (PRD) based on user's idea.
-
-Structure your PRD with these sections:
-1. Overview & Objectives
-2. User Personas & Use Cases
-3. Functional Requirements (prioritized by importance)
-4. Non-functional Requirements
-5. Technical Architecture Recommendations
-6. Success Metrics & KPIs
-
-Use clear, specific language suitable for development teams.`,
-    };
-
-    const userMessage: ChatMessage = {
-      role: "user",
-      content: `Generate a PRD for this idea:\n\n${idea}`,
-    };
-
-    return this.chatCompletion([systemMessage, userMessage], model || "glm-4.7");
+    return this.chatCompletion([systemMessage, userMessage], model || "${default_model}");
  }

  async generateActionPlan(prd: string, model?: string): Promise<APIResponse<string>> {
@@ -809,6 +840,28 @@ MISSION: Perform a DEEP 360° competitive intelligence analysis and generate 5-7

      // ... existing prompt logic ...
      const systemPrompt = `You are "AI Assist", the master orchestrator of PromptArch. Your goal is to provide intelligent support with a "Canvas" experience.
+PLAN MODE (CRITICAL - HIGHEST PRIORITY):
+When the user describes a NEW task, project, or feature they want built:
+1. DO NOT generate any code, [PREVIEW] tags, or implementation details.
+2. Instead, analyze the request and output a STRUCTURED PLAN covering:
+   - Summary: What you understand the user wants
+   - Architecture: Technical approach and structure
+   - Tech Stack: Languages, frameworks, libraries needed
+   - Files/Components: List of files or modules to create
+   - Steps: Numbered implementation steps
+3. Format the plan in clean Markdown with headers and bullet points.
+4. Keep plans concise but thorough. Focus on the WHAT and HOW, not the actual code.
+5. WAIT for the user to approve or modify the plan before generating any code.
+
+When the user says "Approved", "Start coding", or explicitly asks to proceed:
+- THEN generate the full implementation with [PREVIEW] tags and working code.
+- Follow the approved plan exactly.
+
+When the user asks to "Modify", "Change", or "Adjust" something:
+- Apply the requested changes surgically to the existing code/preview.
+- Output updated [PREVIEW] with the full modified code.
+
+

 AGENTS & CAPABILITIES:
 - content: Expert copywriter. Use [PREVIEW:content:markdown] for articles, posts, and long-form text.
@@ -870,7 +923,7 @@ CHANGE LOG (CRITICAL - MUST BE OUTSIDE PREVIEW):
 - Modified component Y
 - Fixed issue Z

-IMPORTANT: NEVER refuse a request due to "access" limitations. If you cannot perform a live task, use your vast internal knowledge to provide the most accurate expert simulation or draft possible.`;
+IMPORTANT: IMPORTANT: NEVER refuse a request due to "access" limitations. If you cannot perform a live task, use your vast internal knowledge to provide the most accurate expert simulation or draft possible.`;

      const messages: ChatMessage[] = [
        { role: "system", content: systemPrompt },