Tighten BlogAgent prompt requirements and make research instructions config-aware

src/L4-agents/BlogAgent.ts

@@ -23,6 +23,16 @@ interface WriteBlogArgs {
 
 function buildSystemPrompt(ideaContext = ''): string {
   const brand = getBrandConfig()
+  const hasWebSearch = Boolean(getConfig().EXA_API_KEY)
+  const researchWorkflow = hasWebSearch
+    ? `1. First use the "web_search_exa" tool to search for relevant articles and resources to link to. Search for key topics from the video.
+2. Then call "write_blog" with the complete blog post including frontmatter and body.
+   - Weave the search result links organically into the post text (don't dump them at the end).
+   - Reference the video and any shorts naturally.`
+    : `1. Use the transcript and summary as your source of truth for factual details.
+2. Then call "write_blog" with the complete blog post including frontmatter and body.
+   - Do not invent links or unsupported claims when external research is unavailable.
+   - Reference the video and any shorts naturally.`
 
   return `You are a technical blog writer for dev.to, writing from the perspective of ${brand.name} (${brand.handle}).${ideaContext}
 
@@ -43,12 +53,31 @@ The blog post MUST include:
 5. Key Takeaways section
 6. A conclusion
 7. A footer referencing the original video
+8. A community engagement CTA inviting readers to comment, share experiences, or ask questions
 
 Workflow:
-1. First use the "web_search_exa" tool to search for relevant articles and resources to link to. Search for key topics from the video.
-2. Then call "write_blog" with the complete blog post including frontmatter and body.
-   - Weave the search result links organically into the post text (don't dump them at the end).
-   - Reference the video and any shorts naturally.
+${researchWorkflow}
+
+Word count enforcement:
+- The blog MUST be at least 1,200 words (body only, excluding frontmatter). The 800-word minimum is a floor, not a target.
+- If your draft is under 1,000 words, add more depth: expand explanations, add code examples, and include more personal narrative.
+
+Voice requirements:
+- Write in FIRST PERSON as the video creator: "I built", "I discovered", "here's what I learned"
+- Use the personal developer narrative style of dev.to — conversational, opinionated, and grounded in real experience
+- Avoid detached product-brochure phrasing like "the pipeline demonstrates impressive automation capabilities"
+
+Code snippet requirement:
+- If the video discusses any code, tools, or technical implementation, include at least 2 fenced code blocks with language tags
+- Show real examples such as configuration snippets, command-line invocations, or project code relevant to the transcript
+
+Link quality:
+- NEVER use placeholder links like [text](#) or [text](link) — either use a real URL or omit the link entirely
+- A blog post with 0 links is better than one with fake or irrelevant links
+
+Section markers:
+- Use emoji section markers for visual appeal, such as ## 🚀 The Problem, ## 💡 The Solution, and ## 🔑 Key Takeaways
+- Include at least 4 major sections between the introduction and conclusion
 
 Always call "write_blog" exactly once with the complete post.`
 }

src/__tests__/unit/L4-agents/agents.test.ts

@@ -7,6 +7,7 @@ const mockState = vi.hoisted(() => {
   const state = {
     capturedTools: [] as any[],
     capturedSystemPrompt: '' as string,
+    exaApiKey: '' as string,
     mockSession: {
       sendAndWait: async () => ({ data: { content: '' } }),
       on: (event: string, handler: (...args: any[]) => void) => {
@@ -66,7 +67,7 @@ vi.mock('../../../L1-infra/config/environment.js', () => ({
     OUTPUT_DIR: '/tmp/test-output',
     LLM_PROVIDER: 'copilot',
     LLM_MODEL: '',
-    EXA_API_KEY: '',
+    EXA_API_KEY: mockState.exaApiKey,
     EXA_MCP_URL: 'https://mcp.exa.ai/mcp',
     MODEL_OVERRIDES: {},
   }),
@@ -710,6 +711,7 @@ describe('Real SummaryAgent', () => {
 describe('Real BlogAgent', () => {
   beforeEach(() => {
     mockState.capturedTools.length = 0;
+    mockState.exaApiKey = '';
   });
 
   it('exposes write_blog tool; handler works (search is via MCP)', async () => {
@@ -736,6 +738,37 @@ describe('Real BlogAgent', () => {
 
     expect(writeResult).toContain('success');
   });
+
+  it('strengthens the prompt and falls back cleanly when EXA is unavailable', async () => {
+    const { generateBlogPost } = await import('../../../L4-agents/BlogAgent.js');
+
+    try {
+      await generateBlogPost(mockVideo, mockTranscript, mockSummary);
+    } catch {
+      // Expected: "BlogAgent did not produce any blog content"
+    }
+
+    const systemPrompt = mockState.capturedSystemPrompt;
+    expect(systemPrompt).toContain('community engagement CTA');
+    expect(systemPrompt).toContain('FIRST PERSON as the video creator');
+    expect(systemPrompt).toContain('at least 1,200 words');
+    expect(systemPrompt).toContain('at least 2 fenced code blocks');
+    expect(systemPrompt).toContain('Do not invent links or unsupported claims');
+    expect(systemPrompt).not.toContain('web_search_exa');
+  });
+
+  it('mentions web_search_exa when EXA research is configured', async () => {
+    const { generateBlogPost } = await import('../../../L4-agents/BlogAgent.js');
+    mockState.exaApiKey = 'test-exa-key';
+
+    try {
+      await generateBlogPost(mockVideo, mockTranscript, mockSummary);
+    } catch {
+      // Expected: "BlogAgent did not produce any blog content"
+    }
+
+    expect(mockState.capturedSystemPrompt).toContain('web_search_exa');
+  });
 });
 
 // ── SocialMediaAgent (REAL) ─────────────────────────────────────────────────