diff --git a/.github/prompts/issue-assistant.md b/.github/prompts/issue-assistant.md
index 93a969a..99d8275 100644
--- a/.github/prompts/issue-assistant.md
+++ b/.github/prompts/issue-assistant.md
@@ -12,23 +12,54 @@ Your role is to provide friendly, informative responses to GitHub issues.
 - Uses Python standard library only (no external dependencies)
 - Supports Python 2.7 and 3.x
 
-## Issue Classification
+## Response Format
 
-**IMPORTANT**: You must classify each issue into ONE of these categories:
+**IMPORTANT**: You must return your response as a JSON object with the following structure:
+
+```json
+{
+  "classification": "bug|feature|question",
+  "response": "Your detailed response here..."
+}
+```
+
+### Classification Categories
+
+Classify each issue into ONE of these categories:
 - **bug**: Software defects, errors, crashes, or unexpected behavior
 - **feature**: New feature requests, enhancements, or improvements
 - **question**: Questions about usage, configuration, documentation, or general inquiries
 
-Return your classification at the START of your response using ONE of the following formats:
-```
+### Response Content Guidelines
+
+When writing the "response" field:
+- **Respond in the same language as the issue title and content** (detect and match the user's language - Chinese, English, or other languages)
+- Keep responses concise but helpful
+- Use proper markdown formatting for code blocks and inline code
+- Reference files and documentation using markdown links
+
+### File References
+
+When referencing files in your response, use these formats:
+1. **Documentation files**: Use markdown links that convert `.md` to `.html` URLs
+   - Format: `[doc/providers/aliesa.md](https://ddns.newfuture.cc/doc/providers/aliesa.html)`
+   - Format: `[doc/config/cli.md](https://ddns.newfuture.cc/doc/config/cli.html)`
+2. **Code files**: Use markdown links to GitHub repository
+   - Format: `[ddns/provider/_base.py](https://github.com/NewFuture/DDNS/blob/master/ddns/provider/_base.py)`
+   - Format: `[tests/test_cache.py](https://github.com/NewFuture/DDNS/blob/master/tests/test_cache.py)`
+3. **Code blocks**: Use proper markdown code fences with language identifiers:
+   ````markdown
+   ```python
+   # Your code here
+   ```
+   ````
+4. **Inline code**: Use backticks for commands and code snippets: `` `ddns --debug` `` or `` `variable_name` ``
 
 ## Response Guidelines
 
 When responding to issues:
-1. Start with the classification tag in the exact format above
-2. Be welcoming and professional
-3. **Respond in the same language as the issue title and content** (detect and match the user's language - Chinese, English, or other languages)
-4. Keep responses concise but helpful
+1. Be welcoming and professional
+2. Be specific and actionable in your guidance
 
 ### Response Strategy by Classification
 
diff --git a/.github/workflows/auto-reply-issue.yml b/.github/workflows/auto-reply-issue.yml
index 7820dd1..8508d74 100644
--- a/.github/workflows/auto-reply-issue.yml
+++ b/.github/workflows/auto-reply-issue.yml
@@ -134,6 +134,7 @@ jobs:
                   ],
                   temperature: 1,
                   max_completion_tokens: 2000,
+                  response_format: { type: "json_object" }
                 })
               });
               
@@ -157,21 +158,42 @@ jobs:
                 return;
               }
               
-              const aiResponse = data.choices[0].message.content.trim();
+              let aiResponse = data.choices[0].message.content.trim();
               
-              // Extract classification from response
-              const classificationMatch = aiResponse.match(/\[CLASSIFICATION:\s*(bug|feature|question)\]/i);
+              // Try to parse JSON response
               let classification = null;
               let commentBody = aiResponse;
               
-              if (classificationMatch) {
-                classification = classificationMatch[1].toLowerCase();
-                // Remove classification tag from the comment body
-                commentBody = aiResponse.replace(/\[CLASSIFICATION:\s*(bug|feature|question)\]\s*/i, '').trim();
-                console.log(`Issue classified as: ${classification}`);
+              try {
+                // Remove markdown code fences if present
+                let jsonContent = aiResponse;
+                if (aiResponse.startsWith('```json') || aiResponse.startsWith('```')) {
+                  jsonContent = aiResponse.replace(/^```(?:json)?\s*\n/, '').replace(/\n```\s*$/, '').trim();
+                }
+                
+                const parsedResponse = JSON.parse(jsonContent);
+                
+                if (parsedResponse.classification && parsedResponse.response) {
+                  classification = parsedResponse.classification.toLowerCase();
+                  commentBody = parsedResponse.response;
+                  console.log(`Issue classified as: ${classification} (JSON format)`);
+                } else {
+                  console.log('JSON response missing required fields, falling back to text format');
+                }
+              } catch (jsonError) {
+                console.log(`Failed to parse JSON response: ${jsonError.message}`);
+                console.log('Falling back to legacy text format extraction');
+                
+                // Fallback: Try legacy format with classification tag
+                const classificationMatch = aiResponse.match(/\[CLASSIFICATION:\s*(bug|feature|question)\]/i);
+                if (classificationMatch) {
+                  classification = classificationMatch[1].toLowerCase();
+                  commentBody = aiResponse.replace(/\[CLASSIFICATION:\s*(bug|feature|question)\]\s*/i, '').trim();
+                  console.log(`Issue classified as: ${classification} (legacy format)`);
+                }
               }
               
-              // Post the response as a comment (without classification tag)
+              // Post the response as a comment
               await github.rest.issues.createComment({
                 owner: context.repo.owner,
                 repo: context.repo.repo,