Blog-Agent-Worker (Pulse)

Source: Blog-Agent-Worker project files, MEMORY.md project notes, Railway deployment

Pipeline diagram (2026-05-09)

BAW pipeline architecture

Single-image overview of the entire system — two entry points (Web Dashboard + Claude Code CLI), 6-column flow (Input → Orchestrator → Content Pipeline → Derivatives → Storage → Export/Live), side modules band (Ad Intelligence, Content Health, Competitive Analysis, SEO+GSC, Lead Magnet, QR), quality+testing+ops band underneath. The two checkpoints — Research Gate (amber) and Final Quality Gate (rose, 141-point validator) — are visible in the third column. Source SVG: baw-pipeline-2026-05-09.svg. Companion image lives in the BAW repo at docs/diagrams/baw-pipeline-2026-05-09.svg + @2x.png.

A cloud-deployable web application for automated marketing content generation using a multi-agent pipeline. Pulse handles blogs, social posts, email sequences, ad intelligence, and competitive analysis through 12 specialized agent definitions. The flagship feature is the 7-agent sequential pipeline (Research, Write, SEO, Edit, Social/Email parallel) which consistently outperforms single-pass generation in quality tests. Deployed on Railway at blog-agent-worker-production.up.railway.app.

What It Does

  • Generates long-form blog content (1500-2500 words) with SEO optimization
  • Runs 7-agent sequential pipeline for highest quality output
  • Validates content against 117-point quality checklist
  • Produces derivative content: social posts, email sequences, ad creative
  • Extracts ad intelligence from Meta, TikTok, Google, LinkedIn ad libraries
  • Scores content health across 5+ data sources

Stack

  • Runtime: Node.js 20
  • Framework: Next.js 15 with Tailwind CSS
  • Database: SQLite (content storage, generation history)
  • AI: Claude Opus 4.6 (orchestrator, writer), Sonnet (SEO, editor, researcher), Haiku (social, email)
  • Deployment: Railway (blog-agent-worker-production.up.railway.app)

Two Generation Pipelines

Standard Pipeline (Single-Pass)

  • Single Claude call generates complete content
  • Faster but lower quality
  • Suitable for high-volume derivative content

Multi-Agent Pipeline (7 Specialized Agents)

Sequential execution with handoffs:

  1. Research Agent (Sonnet, 90s timeout) — Topic research, source gathering, competitive analysis
  2. Writer Agent (Opus, 360s timeout) — Long-form content generation from research brief
  3. SEO Agent (Sonnet, 180s timeout) — Title optimization, meta descriptions, keyword integration, schema
  4. Editor Agent (Sonnet, 180s timeout) — Voice consistency, readability, fact-checking, grammar
  5. Social Agent (Haiku, 60s timeout) — Platform-specific social post derivatives (parallel with Email)
  6. Email Agent (Haiku, 60s timeout) — Email sequence derivatives from blog content (parallel with Social)

The Orchestrator (Opus, 30s timeout) coordinates handoffs and manages pipeline state.

12 Agent Definitions

AgentPurpose
topic-researcherDiscover topics, gaps, trending angles
content-writerGenerate long-form blog content
seo-specialistOptimize for search engines and AI Overviews
content-editorPolish voice, readability, accuracy
social-media-creatorPlatform-specific social posts
email-creatorEmail sequence generation
video-creatorVideo script outlines from content
podcast-creatorPodcast episode scripts
carousel-creatorInstagram/LinkedIn carousel content
thread-creatorTwitter/X thread breakdowns
ad-creative-creatorAd copy and creative briefs
case-study-creatorCase study generation with compliance

Agent Model Allocation

Strategic model selection based on task complexity:

  • Opus — Orchestrator (coordination), Writer (creative long-form). High cost, highest quality
  • Sonnet — Researcher (analysis), SEO (technical optimization), Editor (review). Balanced cost/quality
  • Haiku — Social (derivatives), Email (derivatives). Low cost, high volume

Content Quality Standards

  • Word count: 1500-2500 words (standard), higher for pillar pages
  • Sources: 3+ cited sources per article
  • SEO title: 50-60 characters
  • Meta description: 150-160 characters
  • Answer Capsule: 2-3 sentences after H1 (optimized for AI Overviews citation — the GSC engine can identify which queries need this)
  • FAQ section: 5-7 Q&As with schema markup
  • Reading level: 8th grade (Flesch-Kincaid)
  • Fact density: ~5 statistics per 1500 words with citations, 2024-2026 sources

117-Point Quality Validation

Automated quality gate before human review:

  • Standard content: Must pass 75% of checks (88/117)
  • Pillar pages: Must pass 80% of checks (94/117)
  • Categories: Structure, SEO elements, readability, compliance, sources, formatting, schema, accessibility

Dental Compliance Rules

  • No result guarantees (“We guarantee whiter teeth” is banned)
  • Case study disclaimers required (“Results may vary”)
  • HIPAA awareness (no patient-identifiable information in examples)
  • Dental board advertising rules (state-specific restrictions)

Ad Intelligence Module

Playwright-based extraction from ad libraries:

  • Platforms: Meta Ad Library, TikTok Creative Center, Google Ads Transparency Center, LinkedIn Ad Library
  • 6 Analyzers:
    1. Copy analyzer — Messaging patterns, tone, word choice
    2. Campaign analyzer — Budget allocation, targeting, scheduling
    3. Case-study analyzer — Before/after patterns, proof points
    4. Pain-categorizer — Customer pain points by category
    5. Audience analyzer — Demographic and psychographic targeting
    6. Swipe-file builder — Best-performing creative for reference

Content Health Scoring

Multi-source health assessment formula: (rawSum / maxAchievable) * 100

Data SourceWhat It Measures
FreshnessDate decay — older content scores lower
CWV (PageSpeed)Core Web Vitals performance
GA4Engagement metrics (time on page, scroll depth)
GSCClick-through rate and position data
Link healthInternal/external link status
Geo readinessLocal SEO and AI Overviews optimization

No single source is sufficient. The composite score reveals content that looks good on one metric but fails on others.

Key Takeaways

  • Multi-agent sequential beats single-pass generation every time — the research phase alone dramatically improves factual accuracy
  • Answer Capsule + FAQ are now required for GEO (Generative Engine Optimization / AI Overviews citation)
  • Model allocation matters: Opus for creative, Sonnet for analytical, Haiku for high-volume derivatives saves 60%+ on API costs vs using Opus for everything — see SEO Patterns Learned for cross-project model selection guidance
  • 117-point quality validation catches issues before human review, reducing revision cycles
  • Content health scoring from 5+ data sources reveals problems that single-metric tools miss
  • Ad intelligence provides competitive creative analysis that informs content strategy

Try It

  1. Visit blog-agent-worker-production.up.railway.app for the web interface
  2. Try the multi-agent pipeline on a topic: select “Multi-Agent” mode, enter a dental marketing topic
  3. Review the 117-point quality report that accompanies each generation
  4. Use the ad intelligence module to analyze competitor creative before writing competitive content
  5. Check content health scores for existing blog posts to identify refresh candidates