The Problem: Writing a Blog Post Took Me 8 Hours
I love writing. But the process of creating a single high-quality blog post was killing me. Research: 2 hours. Outlining: 1 hour. Writing: 3 hours. SEO optimization: 1 hour. Finding images and formatting: 1 hour. That's 8 hours minimum for one post. And I want to publish 3-4 times per week across three verticals — fitness, finance, and technology.
The math didn't work. I'm one person. I don't have a content team. I don't have an editor or an SEO specialist or a designer. I have me, my laptop, and a deep belief that AI can handle the parts of this process that don't require my personal voice.
So I built a system. Six AI agents, each with a specific role, working in sequence to take a raw idea and turn it into a fully structured, SEO-optimized, ready-to-publish blog post. I built the first version in 3 days.
The Architecture: 6 Agents, One Pipeline
Think of it like an assembly line. Each agent has one job. Each agent's output becomes the next agent's input. The pipeline flows in one direction: idea → research → strategy → writing → SEO → image.
Agent 0.5 — The Idea Scout
Model: Claude Sonnet (with web search)
Job: Find content ideas worth writing about.
This agent scans trends, analyzes what's performing in the fitness/finance/tech space in India, identifies content gaps, and generates ideas scored on virality potential, audience need, and alignment with the platform's voice. It runs weekly and produces a ranked list of 15-20 ideas.
Each idea comes with: a proposed title, the target vertical, the content type (knowledge post or letter), a virality score, and a one-paragraph justification for why this topic matters now.
Agent 0 — The Researcher
Model: Claude Sonnet
Job: Deep research on the approved idea.
Once I approve an idea from the Scout's list, the Researcher takes over. It gathers data points, statistics, expert opinions, counterarguments, and relevant studies. For a fitness post, it might pull recent research on muscle protein synthesis. For a finance post, it might grab the latest SEBI data on mutual fund flows.
Output: a structured research brief with key facts, data points, sources, and suggested angles.
Agent 1 — The Strategist
Model: Claude Sonnet
Job: Build the content architecture.
The Strategist takes the research brief and designs the article structure. It determines: the hook (how the article opens), the sections and their order, where personal stories should be inserted, where data should appear, what the emotional arc should be, and what the call to action is.
This agent also specifies tone guidelines for the Writer agent — "open with vulnerability, transition to authority, end with challenge to the reader."
Output: a detailed outline with section headers, talking points per section, emotional beats, and structural notes.
Agent 2 — The Writer
Model: Claude Opus
Job: Write the actual article.
This is the only agent that uses Opus — the most capable model — because writing quality is everything. This agent takes the Strategist's outline and produces the full article in HTML. It follows the voice rules strictly: personal stories first, direct language, no fluff, specific numbers, tough love where needed.
I will never downgrade this agent to a smaller model. The writing is the product. The writing is what people read, share, and remember. Cutting cost on the Writer is cutting quality on the only thing that matters.
Output: complete content_html ready for the blog template.
Agent 3 — The SEO Optimizer
Model: Claude Sonnet
Job: Optimize for search without destroying the voice.
This is the delicate agent. It takes the Writer's output and adds SEO elements: meta title, meta description, focus keyword, secondary keywords, internal link suggestions, and minor content adjustments for keyword density. The critical rule: it cannot change the voice or add generic SEO phrases. "In this comprehensive guide" — absolutely not. The optimization must be invisible to the reader.
Output: SEO metadata + slightly optimized content_html.
Vision Agent — The Image Mind
Model: Claude Sonnet (vision)
Job: Suggest and evaluate images.
This agent analyzes the article content, suggests what type of images would work (personal photos, infographics, screenshots), evaluates uploaded photos for quality and relevance, and writes alt text and captions. It connects with the fal.ai pipeline for image enhancement when needed.
Output: image recommendations, alt text, placement suggestions.
How It All Connects — The Pipeline Flow
Here's the complete flow visualized as text:
IDEA SCOUT → [I approve an idea] → RESEARCHER → STRATEGIST → WRITER → SEO OPTIMIZER → VISION AGENT → [I review and publish]
Two human checkpoints. I approve the idea at the start. I review the final output at the end. Everything in between is automated. The entire pipeline — from approved idea to publishable post — takes about 4-6 minutes of compute time.
Compare that to 8 hours of manual work.
The Tech Stack Behind It
- AI Models: Anthropic Claude API — Sonnet for most agents, Opus for the Writer. All model assignments stored in a settings table in Supabase, configurable from the admin panel without code changes.
- Orchestration: Next.js API routes handle the pipeline flow. Each agent is a separate function call with its own system prompt and context window.
- Database: Supabase (PostgreSQL). Every pipeline run is logged — the inputs, outputs, model used, tokens consumed, and time taken. This data helps me optimize prompts and identify bottlenecks.
- Frontend: Admin dashboard in Next.js shows the pipeline stages in real-time. I can see each agent's output, edit it if needed, and push to the next stage or publish directly.
- Image Pipeline: fal.ai for image enhancement. Photos uploaded via Telegram bot or admin panel get analyzed and enhanced automatically.
Why This Matters Beyond My Blog
This architecture isn't specific to blog writing. The pattern — specialized agents in a pipeline, each with a narrow job, human approval gates at key points — applies to any content creation workflow. Marketing teams, newsrooms, educational platforms — anyone producing content at scale can use this pattern.
The key insight: don't build one AI that does everything. Build multiple agents that each do one thing extremely well, and chain them together. Specialization produces quality. Generalization produces mediocrity.
"I didn't replace myself with AI. I gave myself a team of six that works at the speed of thought."
The blog engine is live. It powers the content you're reading right now. Every post on icanbefitter.com flows through this pipeline — though I still write the Letters to Avyaansh myself, by hand. Some things shouldn't be automated.
Full code walkthrough and prompt engineering details coming in future posts. Follow along if you want to build something similar.