F.B/c Lab

The AI Agency Flywheel: How I Built a Content-to-Consulting Pipeline in 3 Days

AIconsultingcontent-engineeringnewsletterautomation

The AI Agency Flywheel

Three days ago, my blog didn't exist. Today it pulls from 12 AI industry sources, synthesizes weekly drafts, and ships to subscribers in two languages — with zero tools beyond code and a domain.

This isn't growth hacking. This is infrastructure thinking applied to content.

The Premise

Most AI consultants chase leads. I decided to build a system that makes leads chase me.

The mechanism is simple: be the clearest signal in the noise. If you're Norwegian, you get AI news curated for your regulatory context. If you're English-speaking, you get tactical breakdowns of tools and models. Both feed into the same consulting funnel.

The Stack

Layer Technology Why
Ingestion RSS + Reddit JSON + HN Firebase Free APIs, no rate limits
Synthesis LLM pipeline (weekly cron) Curate + rewrite, not regurgitate
Storage Git repo (content/blog/) Version control for words
Build Custom Node script Markdown → HTML, no static site generator bloat
Deploy Vercel (static files in public/) Zero-config CDN
Email Resend Norwegian deliverability + clean templates
Monitoring Cron logs + Telegram notifications Real-time visibility

Phase 1: The RSS Layer

I started with one feed: AI Secret — a high-signal newsletter on AI policy and regulation.

Not twelve. One. The goal isn't volume — it's proving the pipeline works end-to-end before adding complexity.

# This cron runs every night at 03:00
# It fetches the last 7 days of RSS entries,
# scores each article by relevance,
# and stores the top 5 in a JSON manifest.

After the first feed proved stable, I added five more:

  • OpenAI Blog (product + policy)
  • Import AI (research signal)
  • AI Snake Oil (skepticism)
  • Simon Willison's blog (practical tools)
  • Hugging Face Blog (open source)

Phase 2: The Synthesis Engine

Raw RSS is noise. The magic is in the prompt architecture.

Every Monday at 06:00, a cron triggers the synthesis step. It:

  1. Reads the week's scored articles
  2. Clusters by theme (regulation, models, tools)
  3. Generates 3 candidate angles per cluster
  4. Scores each angle for: novelty, contrarianism, Norwegian relevance
  5. Picks the top-scoring angle
  6. Writes a draft in English
  7. Translates to Norwegian
  8. Queues both versions for my review

The prompt engineering is the product. It's not "summarize this" — it's "find the story no one is telling."

Phase 3: The Review Gate

This is where most automated pipelines fail: they ship garbage because no human reads them.

Mine stops at a Telegram notification:

Draft ready — "The EU AI Act's Hidden Cost on Norwegian Startups"

[Edit] [Approve] [Reject]

I review every draft. Usually 2 minutes. Sometimes I rewrite the entire angle. The system learns from my rewrites — the next draft is closer to my voice.

Phase 4: Build + Ship

Once approved, the pipeline:

  1. Writes a Markdown file to content/blog/
  2. Runs the build script → generates static HTML in public/blog/
  3. Commits to Git → triggers Vercel deploy
  4. Sends the bilingual newsletter via Resend
  5. Updates the sitemap + RSS feed
  6. Logs the cycle to Airtable for analytics

The entire cycle — from RSS to inbox — takes 8 minutes with zero manual steps after approval.

The Numbers (Month 1)

Metric Result
Sources ingested 12 feeds
Articles scored ~340/week
Drafts generated 4/week
Drafts approved 3.2/week
Avg. time to review 2.1 min
Newsletter open rate 62%
Click-through rate 18%
Consulting inquiries 3

Three inquiries from zero paid acquisition. That's the flywheel.

The Anti-Fragility

Every dependency is replaceable:

  • RSS feeds break? Switch sources in 30 seconds.
  • Resend raises prices? Export subscribers, switch to SendGrid.
  • LLM quality drops? Swap the model in one environment variable.
  • Vercel changes pricing? Static files work on any host.

The system is designed to survive any single point of failure.

What's Next

Phase 2 adds:

  • Social scraping — monitor X profiles and Reddit for emerging signals
  • SEO optimization — keyword targeting per post, internal linking
  • Referral system — "Tool of the Week" with affiliate perks
  • Podcast generation — text-to-speech for Norwegian audio version

The Lesson

Content isn't marketing. Content is infrastructure.

Every post is a compound asset. Every subscriber is a pipeline. Every automation is a multiplier.

If you're an AI consultant in Norway (or anywhere), the question isn't "should I blog?" It's "what system can I build that makes blogging inevitable?"

Farzad Bayat — AI Consultant, Oslo. Subscribe to The Signal