HashBuilds Secure Prompts

Instructions

Implementation Reference

• src/index.ts:221-298 (handler)

  The core handler function that takes a list of detected prompts from the codebase, analyzes their context (user-facing vs internal) using patterns and file paths, assigns confidence scores and suggested actions, computes statistics, and returns a structured AuditResult with recommendations for registration.
  async function auditPrompts(args: { prompts: Array<{ filePath: string; lineNumber: number; promptText: string; surroundingCode?: string; }>; }): Promise<AuditResult> { const candidates: PromptCandidate[] = []; for (const prompt of args.prompts) { const surrounding = prompt.surroundingCode || ''; const filePath = prompt.filePath; // Determine context based on file path and surrounding code let context: 'user_facing' | 'internal' | 'unknown' = 'unknown'; let confidence = 50; // Check file path patterns const isPublicFile = /public\/|PROMPT_.*\.txt/i.test(filePath); const isApiFile = /api\/|\.server\.|route\.ts/i.test(filePath); const isComponentFile = /components?\/|\.tsx$/i.test(filePath); // Check surrounding code patterns const hasUserFacingIndicators = USER_FACING_PATTERNS.some(p => p.test(surrounding) || p.test(filePath)); const hasInternalIndicators = INTERNAL_PATTERNS.some(p => p.test(surrounding) || p.test(filePath)); // Determine context if (isPublicFile || hasUserFacingIndicators) { context = 'user_facing'; confidence = isPublicFile ? 95 : 75; } else if (isApiFile || hasInternalIndicators) { context = 'internal'; confidence = isApiFile ? 90 : 70; } else if (isComponentFile) { // Components could be either - need review context = hasUserFacingIndicators ? 'user_facing' : 'unknown'; confidence = 60; } // Determine suggested action let suggestedAction: 'register_badge' | 'audit_only' | 'review' = 'review'; if (context === 'user_facing' && confidence >= 70) { suggestedAction = 'register_badge'; } else if (context === 'internal' && confidence >= 70) { suggestedAction = 'audit_only'; } // Create preview (first 100 chars) const preview = prompt.promptText.substring(0, 100) + (prompt.promptText.length > 100 ? '...' : ''); candidates.push({ filePath: prompt.filePath, lineNumber: prompt.lineNumber, promptText: prompt.promptText, context, confidence, suggestedAction, preview, }); } // Calculate summary stats const userFacing = candidates.filter(c => c.context === 'user_facing').length; const internal = candidates.filter(c => c.context === 'internal').length; const needsReview = candidates.filter(c => c.context === 'unknown' || c.confidence < 70).length; const summary = `Found ${candidates.length} prompts: ${userFacing} user-facing (recommend badges), ${internal} internal (audit only), ${needsReview} need manual review.`; return { totalFound: candidates.length, userFacing, internal, needsReview, prompts: candidates, summary, }; }
• src/index.ts:473-504 (schema)

  JSON schema defining the input parameters for the audit_prompts tool: requires a 'prompts' array, each item with filePath, lineNumber, promptText, and optional surroundingCode.
  inputSchema: { type: "object", properties: { prompts: { type: "array", description: "Array of prompts found in the codebase", items: { type: "object", properties: { filePath: { type: "string", description: "Path to the file containing the prompt", }, lineNumber: { type: "number", description: "Line number where the prompt starts", }, promptText: { type: "string", description: "The full prompt text", }, surroundingCode: { type: "string", description: "Optional: Code around the prompt (helps determine if user-facing)", }, }, required: ["filePath", "lineNumber", "promptText"], }, }, }, required: ["prompts"], },
• src/index.ts:455-505 (registration)

  Registers the audit_prompts tool in the MCP server's listTools response, including name, detailed description on usage for codebase scanning, and input schema.
  { name: "audit_prompts", description: "Analyze a list of prompts found in a codebase and categorize them as user-facing (needs badge) " + "or internal (audit only). This tool helps users who already have prompts in their codebase " + "understand which ones should be registered with secure badges vs which are internal-only.\n\n" + "HOW TO USE:\n" + "1. First, search the codebase for prompts using patterns like:\n" + " - Files matching: public/PROMPT_*.txt, **/prompt*.ts\n" + " - Code patterns: 'You are a', 'systemPrompt', 'SYSTEM_PROMPT', role: 'system'\n" + "2. Extract the prompt text and file location for each found prompt\n" + "3. Call this tool with the prompts array\n" + "4. Present the audit results to the user, showing:\n" + " - User-facing prompts that should get security badges\n" + " - Internal prompts that are safe but should be audited\n" + " - Prompts needing manual review\n" + "5. Ask the user which prompts they want to register for badges\n" + "6. Use register_secure_prompt for each selected prompt", inputSchema: { type: "object", properties: { prompts: { type: "array", description: "Array of prompts found in the codebase", items: { type: "object", properties: { filePath: { type: "string", description: "Path to the file containing the prompt", }, lineNumber: { type: "number", description: "Line number where the prompt starts", }, promptText: { type: "string", description: "The full prompt text", }, surroundingCode: { type: "string", description: "Optional: Code around the prompt (helps determine if user-facing)", }, }, required: ["filePath", "lineNumber", "promptText"], }, }, }, required: ["prompts"], }, },
• src/index.ts:573-647 (registration)

  Handles incoming CallToolRequest for 'audit_prompts' in the MCP server: validates arguments, calls the auditPrompts handler, enhances result with guidance text, and returns formatted response.
  case "audit_prompts": { const typedArgs = args as { prompts: Array<{ filePath: string; lineNumber: number; promptText: string; surroundingCode?: string; }>; }; if (!typedArgs.prompts || !Array.isArray(typedArgs.prompts)) { throw new McpError(ErrorCode.InvalidParams, "prompts array is required"); } if (typedArgs.prompts.length === 0) { return { content: [ { type: "text", text: JSON.stringify({ totalFound: 0, userFacing: 0, internal: 0, needsReview: 0, prompts: [], summary: "No prompts provided for analysis.", }, null, 2), }, ], }; } const result = await auditPrompts(typedArgs); // Add guidance for the AI agent const guidance = ` ## Prompt Audit Results ${result.summary} ### Recommended Actions: **User-Facing Prompts (${result.userFacing}):** These prompts are likely displayed to users or have copy buttons. Consider registering them with secure badges: ${result.prompts.filter(p => p.context === 'user_facing').map(p => `- ${p.filePath}:${p.lineNumber} - "${p.preview}" (${p.confidence}% confidence)` ).join('\n') || '- None found'} **Internal Prompts (${result.internal}):** These appear to be backend/API prompts. They should be secure but don't need public badges: ${result.prompts.filter(p => p.context === 'internal').map(p => `- ${p.filePath}:${p.lineNumber} - "${p.preview}"` ).join('\n') || '- None found'} **Needs Review (${result.needsReview}):** These prompts need manual review to determine if they're user-facing: ${result.prompts.filter(p => p.context === 'unknown' || p.confidence < 70).map(p => `- ${p.filePath}:${p.lineNumber} - "${p.preview}" (${p.confidence}% confidence)` ).join('\n') || '- None found'} ### Next Steps: 1. Ask the user: "Would you like to register any of these prompts with security badges?" 2. For user-facing prompts, use register_secure_prompt 3. Show them the badge options (full badge, compact link, icon button) `; return { content: [ { type: "text", text: JSON.stringify({ ...result, guidance }, null, 2), }, ], }; }
• src/index.ts:182-215 (helper)

  Regular expression pattern arrays used by the handler to detect prompt-like strings and classify them as user-facing or internal based on surrounding code and file paths.
  const PROMPT_PATTERNS = [ // System prompts /system\s*:\s*[`"']([^`"']{50,})[`"']/gi, /systemPrompt\s*[:=]\s*[`"']([^`"']{50,})[`"']/gi, /SYSTEM_PROMPT\s*=\s*[`"']([^`"']{50,})[`"']/gi, // Role definitions /["'`]You are (?:a |an )?[^"'`]{30,}["'`]/gi, /role\s*:\s*["'`]system["'`][\s\S]{0,50}content\s*:\s*["'`]([^"'`]{50,})["'`]/gi, // Prompt variables /(?:const|let|var)\s+\w*[Pp]rompt\w*\s*=\s*[`"']([^`"']{50,})[`"']/gi, // Template literals with instructions /`[^`]*(?:instructions?|guidelines?|rules?)[^`]*`/gi, ]; // Patterns suggesting user-facing (copy button, displayed to user) const USER_FACING_PATTERNS = [ /copy.*button|copyable|clipboard/i, /onClick.*copy|handleCopy/i, /data-prompt|promptText.*prop/i, /user.*can.*copy|copy.*to.*clipboard/i, /<pre>|<code>|CodeBlock/i, /PROMPT_.*\.txt|public\/.*prompt/i, ]; // Patterns suggesting internal-only const INTERNAL_PATTERNS = [ /api\/|server|backend/i, /process\.env|getServerSide/i, /internal|private|system/i, /\.server\.|route\.ts|api\//i, ];