Hi-AI

// Enhance prompts using Gemini API prompting strategies import { ToolResult, ToolDefinition } from '../../types/tool.js'; export const enhancePromptGeminiDefinition: ToolDefinition = { name: 'enhance_prompt_gemini', description: 'prompt improvement|gemini strategies|quality improvement|prompt enhancement|gemini strategies|quality improvement - Enhance prompts using Gemini API prompting strategies (Few-Shot, Output Format, Context)', inputSchema: { type: 'object', properties: { prompt: { type: 'string', description: 'The original prompt to enhance' }, agent_role: { type: 'string', description: 'The role of the agent that will receive this prompt (e.g., "Specification Agent", "Planning Agent")' }, strategies: { type: 'array', items: { type: 'string', enum: ['few-shot', 'output-format', 'context-placement', 'decomposition', 'parameters'] }, description: 'Specific Gemini strategies to apply. If not provided, all strategies will be applied.' } }, required: ['prompt'] }, annotations: { title: 'Enhance Prompt (Gemini Strategies)', audience: ['user', 'assistant'] } }; interface PromptEnhancement { strategy: string; description: string; applied: string; improvement: string; } export async function enhancePromptGemini(args: { prompt: string; agent_role?: string; strategies?: string[]; }): Promise<ToolResult> { const { prompt, agent_role, strategies } = args; const allStrategies = ['few-shot', 'output-format', 'context-placement', 'decomposition', 'parameters']; const strategiesToApply = strategies && strategies.length > 0 ? strategies : allStrategies; const enhancements: PromptEnhancement[] = []; // 1. Few-Shot Examples if (strategiesToApply.includes('few-shot')) { enhancements.push({ strategy: 'Add Few-Shot Examples', description: 'Add 2-3 high-quality examples to guide the model to learn patterns', applied: addFewShotExamples(prompt, agent_role), improvement: 'Improve consistency by clarifying format, expression, and scope' }); } // 2. Output Format Specification if (strategiesToApply.includes('output-format')) { enhancements.push({ strategy: 'Output Format Specification', description: 'Specify structured format with XML tags or markdown headers', applied: specifyOutputFormat(prompt, agent_role), improvement: 'Clarify desired response structure for easier parsing' }); } // 3. Context Placement if (strategiesToApply.includes('context-placement')) { enhancements.push({ strategy: 'Context Placement Optimization', description: 'Place long context before specific requests (Gemini 3 optimization)', applied: optimizeContextPlacement(prompt), improvement: 'Help model utilize context more effectively' }); } // 4. Prompt Decomposition if (strategiesToApply.includes('decomposition')) { enhancements.push({ strategy: 'Prompt Decomposition', description: 'Decompose complex tasks into multiple steps and chain them', applied: decomposePrompt(prompt, agent_role), improvement: 'Improve output quality at each step, easier debugging' }); } // 5. Parameter Tuning Suggestions if (strategiesToApply.includes('parameters')) { enhancements.push({ strategy: 'Parameter Tuning Suggestions', description: 'Suggest optimal values for Temperature, Top-K, Top-P, Max Tokens', applied: suggestParameters(prompt, agent_role), improvement: 'Optimize model behavior for task type' }); } const result = { original_prompt: prompt, agent_role: agent_role || 'Generic Agent', strategies_applied: strategiesToApply, enhancements, enhanced_prompt: combineEnhancements(prompt, enhancements), summary: generateSummary(enhancements) }; const output = formatOutput(result); return { content: [{ type: 'text', text: output }] }; } // Helper methods function addFewShotExamples(prompt: string, agent_role?: string): string { const examples = { 'Specification Agent': ` **Example 1: Push Notification Settings** Input: "Let me turn on/off notifications for comments, likes, and follows" Output: - Notification types: 6 (comments, likes, follows, announcements, events, marketing) - Setting method: ON/OFF toggle - Design reference: iOS Settings > Notifications - Tech stack: FCM (currently in use) **Example 2: User Profile Editing** Input: "Let me change my profile picture and bio" Output: - Editable fields: Profile picture, bio, display name - Validation: Image size < 5MB, bio < 500 chars - UI pattern: Inline editing (not modal) - Save method: Auto-save (debounce 500ms)`, 'Planning Agent': ` **Example 1: API Endpoint Addition** Input: "User follow/unfollow feature" Output: - Phase 1: Backend (8 hours) - DB schema (follows table) - API endpoints (POST /follows, DELETE /follows/:id) - Business logic (duplicate prevention, self-follow prevention) - Phase 2: Frontend (6 hours) - Follow button component - Followers/following list - Cost impact: +$0 (using existing infrastructure) **Example 2: Real-time Notifications** Input: "Real-time notification when someone comments" Output: - Phase 1: WebSocket server setup (12 hours) - Phase 2: Client subscription logic (8 hours) - Phase 3: Notification UI (6 hours) - Cost impact: +$20/month (Redis Pub/Sub)` }; return examples[agent_role as keyof typeof examples] || ` **Add Few-Shot examples matching the task type**: - Example 1: [Specific input] → [Clear output] - Example 2: [Different input format] → [Consistent output format] - Example 3: [Edge case] → [How to handle] **Guidelines**: - 2-3 examples (avoid overfitting) - Include diverse scenarios - Maintain consistent output format`; } function specifyOutputFormat(prompt: string, agent_role?: string): string { const formats = { 'Specification Agent': ` **Output Format (Markdown + YAML frontmatter)**: \`\`\`markdown --- title: [Feature Name] priority: [HIGH/MEDIUM/LOW] created: [Date] --- # SPEC: [Feature Name] ## REQ-001: [Requirement Title] **WHEN** [Condition] **THEN** [Result] ### Acceptance Criteria - [ ] [Criterion 1] - [ ] [Criterion 2] \`\`\` **Response prefix**: Start with "# SPEC: " to guide model to complete in correct format`, 'Planning Agent': ` **Output Format (Structured Markdown)**: \`\`\`markdown # PLAN: [Feature Name] ## Architecture - Backend: [Tech Stack] - Frontend: [Tech Stack] - Database: [Schema Changes] ## Timeline | Phase | Tasks | Duration | |-------|-------|----------| | 1 | ... | 8h | ## Cost Analysis - Infrastructure: +$X/month - Third-party: +$Y/month - Total: +$Z/month \`\`\` **Response prefix**: Start with "# PLAN: "` }; return formats[agent_role as keyof typeof formats] || ` **Output Format Specification**: - Separate sections with markdown headers (##, ###) - Use tables, bullet points, checkboxes - Optionally label semantic components with XML tags e.g.: <analysis>...</analysis>, <recommendation>...</recommendation> - Use response prefix to guide completion e.g.: Starting with "Analysis result: " makes model complete with analysis content`; } function optimizeContextPlacement(prompt: string): string { return ` **Context Placement Optimization (Gemini 3 Recommended)**: 1. **Place long context first**: \`\`\` [Tech stack info (CLAUDE.md content)] [Existing codebase structure] [Related SPEC/PLAN documents] --- Specific request below --- [User's specific question or task] \`\`\` 2. **Structure context**: - Group by category (tech stack, architecture, business logic) - Emphasize important constraints repeatedly - Label clearly for reference 3. **Place explicit instructions**: - Specific task instructions after context - Step-by-step guidelines (1, 2, 3...) - Output format examples`; } function decomposePrompt(prompt: string, agent_role?: string): string { const isComplex = prompt.length > 200 || prompt.includes('and') || prompt.includes('also') || prompt.includes('and'); if (!isComplex) { return '**Prompt decomposition not needed**: Simple task, single prompt is sufficient.'; } return ` **Prompt Decomposition (Sequential Chain)**: **Step 1: Information Gathering** \`\`\` Prompt: Identify required information for "${prompt.slice(0, 100)}..." Output: List of required tech stack, constraints, prerequisite tasks \`\`\` **Step 2: Analysis and Planning** \`\`\` Prompt: Create implementation plan based on Step 1 information. Input: [Step 1 output] Output: Phase-by-phase tasks, timeline, risks \`\`\` **Step 3: Detailed Task Generation** \`\`\` Prompt: Decompose Step 2 plan into actionable tasks. Input: [Step 2 output] Output: Task list (with dependencies) \`\`\` **Benefits**: - Improved output quality at each step - Intermediate validation possible - Re-run only specific step on error`; } function suggestParameters(prompt: string, agent_role?: string): string { const isCreative = prompt.includes('design') || prompt.includes('idea') || prompt.includes('suggest'); const isDeterministic = prompt.includes('analyze') || prompt.includes('calculate') || prompt.includes('verify'); let temperature = 1.0; // Gemini 3 default let topP = 0.95; let topK = 40; if (isCreative) { temperature = 1.0; topP = 0.95; } else if (isDeterministic) { temperature = 0.2; topP = 0.8; topK = 20; } return ` **Recommended Parameters** (Based on task characteristics): - **Temperature**: ${temperature} ${temperature > 0.7 ? 'Suitable for creative tasks (explore diverse options)' : 'Suitable for deterministic tasks (consistent output)'} - **Top-P**: ${topP} Select tokens up to ${topP * 100}% cumulative probability - **Top-K**: ${topK} Consider only top ${topK} tokens - **Max Output Tokens**: ${agent_role === 'Specification Agent' ? '4000 (detailed document)' : '2000 (general)'} - **Stop Sequences**: ["---END---", "\`\`\`"] (optional) **Note**: - Gemini 3 recommends keeping temperature at default 1.0 - Avoid deviating significantly from defaults to prevent unexpected behavior`; } function combineEnhancements(prompt: string, enhancements: PromptEnhancement[]): string { let enhanced = `# Enhanced Prompt\n\n`; enhanced += `## Original Request\n${prompt}\n\n`; enhanced += `---\n\n`; enhancements.forEach((e, i) => { enhanced += `## Enhancement ${i + 1}: ${e.strategy}\n\n`; enhanced += `${e.applied}\n\n`; }); return enhanced; } function generateSummary(enhancements: PromptEnhancement[]): string { return `Applied ${enhancements.length} Gemini prompting strategies to improve prompt quality: ${enhancements.map(e => `- ${e.strategy}: ${e.improvement}`).join('\n')}`; } function formatOutput(result: any): string { let output = `# Gemini Prompt Enhancement\n\n`; output += `**Original Prompt**: ${result.original_prompt}\n`; output += `**Target Agent**: ${result.agent_role}\n`; output += `**Applied Strategies**: ${result.strategies_applied.join(', ')}\n\n`; output += `---\n\n`; result.enhancements.forEach((e: PromptEnhancement, i: number) => { output += `## ${i + 1}. ${e.strategy}\n\n`; output += `**Description**: ${e.description}\n\n`; output += `**Applied Content**:\n${e.applied}\n\n`; output += `**Improvement Effect**: ${e.improvement}\n\n`; output += `---\n\n`; }); output += `## Enhanced Prompt\n\n`; output += '```markdown\n' + result.enhanced_prompt + '\n```\n\n'; output += `## Summary\n\n${result.summary}`; return output; }