Read the full submission
SUMMARY
· Built a production-grade, multi-agent AI system as a non-technical founder using AI Mastery principles
· Reduced LinkedIn content creation time by 75–85% (2–4 hours → 15–30 minutes per post)
· Transformed content creation from a manual bottleneck into scalable, AI-orchestrated infrastructure
· Designed a reusable architecture extendable to newsletter, long-form writing, and book development
WHY I BUILT THIS
I'm in the founding phase of my own business. LinkedIn consistency became non-negotiable — I post 2–3 times per week. High-quality thought leadership posts took 2–4 hours each, even when iterating with ChatGPT, Claude, and Gemini.
The outputs were helpful. The workflow was not. Briefing my thoughts, drafting, crafting hooks, checking brand voice, finetuning text, choosing the right timing — it did not scale.
So instead of using AI to help me write posts, I decided to architect AI to own the process.
Many of my ideas come while walking. So I built the system around that reality: raw ideas dictated on my iPhone get transformed into publish-ready LinkedIn posts by 4 specialized AI agents — fully automated, running 24/7 in n8n Cloud, polling for new ideas every 5 minutes.
ACHIEVEMENTS & MEASURABLE RESULTS
· Content creation time: 2–4 hours → 15–30 minutes per post (75–85% reduction)
· Pipeline processing time: ~2.5 minutes end-to-end (fully automated)
· 13 n8n nodes orchestrated in a conditional pipeline with two execution paths
· 5 integrated systems: n8n Cloud, Anthropic Claude API, Airtable, Slack, GitHub
· 5 production versions in 5 days (V1→V5), driven by real usage
· 9 git commits documenting the full evolution
· Parallel processing: each Airtable record triggers independently — multiple ideas run simultaneously
· System scales to 10+ posts per week without architectural changes
SYSTEM ARCHITECTURE — 4-Agent Pipeline (Claude Sonnet 4.5)
Each idea flows sequentially through a defined chain of responsibility:
1. Draft Writer — transforms a structured briefing into narrative (creative temperature)
2. Hook Generator — produces 3 hook variants optimized for LinkedIn engagement (exploratory temperature)
3. Timing Engine — calculates optimal posting windows using a unique multi-variable model (deterministic temperature)
4. QA Editor — enforces brand voice fidelity, scores compliance, and flags AI-typical language (precise temperature)
Each agent operates with:
- XML-structured system prompts with explicit role definitions
- Defined input/output contracts (each agent knows its upstream source and downstream consumer)
- Calibrated temperature settings matching its function
- Guardrails and output format specifications
This is modular agent orchestration — not a single prompt.
CONDITIONAL ROUTING
Two execution modes via IF node:
- "Generate Draft" → Full AI creation from an idea briefing
- "Review Post" → I write the post myself, AI adds hooks + timing + QA
AI accelerates. I retain final authority.
MOBILE-FIRST INPUT
Ideas captured in real-world context:
- iPhone home screen shortcut → Airtable form
- Wispr Flow voice dictation — I dictate in 30 seconds while walking
- n8n Schedule Trigger polls every 5 minutes — pipeline runs automatically
Before this shortcut, ideas were lost. Now they enter the system in under 30 seconds and trigger a full production pipeline.
THE TIMING ENGINE (Key Differentiator)
The timing agent maps post themes to the Five Elements from Taoist astrology, cross-references lunar phases, and overlays LinkedIn DACH engagement data.
It produces 3–5 concrete publishing windows across 6 weeks.
No other submission will combine thematic mapping, lunar cycles, and platform analytics in a structured decision engine.
QA & BRAND VOICE ENGINEERING
- 23KB XML brand voice definition
- 8 real LinkedIn posts as calibration anchors
- Dedicated Voice Guide document
- QA agent scoring brand voice compliance across 5 dimensions (1–5 scale)
- AI language red-flag detection (flags words like "navigate", "crucial", "landscape")
Voice consistency is enforced structurally — not stylistically guessed.
INTELLIGENT CONTENT HUB (Airtable)
The LinkedIn Posts table is not just storage — it's an AI-enhanced content management system:
- Automated record lifecycle: Airtable captures the idea → n8n processes through all 4 agents → Slack delivers a formatted report → LinkedIn Posts table stores the complete output → source record auto-deletes
- Engagement rate and performance rating calculated via embedded formulas and AI agents
- Content pillar mapping and hook effectiveness tracking
- Creates a feedback loop: I see which topics, hooks, and posting times perform best — and feed that insight into future content ideas
This "content brain" grows with every published post. Whether I'm writing LinkedIn posts, preparing a book, or designing workshops — every piece of content feeds back into the same knowledge base.
Next evolution: automated analytics ingestion via LinkedIn API or CSV detection — closing the loop from idea → draft → publish → performance → next idea.
ERROR RESILIENCE
The system gracefully handles inconsistent AI output formats and malformed JSON — critical for any autonomous pipeline running without human supervision. A dedicated state machine (repairJson) ensures the pipeline continues even when model outputs deviate from expected structure.
TOOLS
· n8n Cloud — workflow orchestration (13 nodes, Schedule Trigger, HTTP Request nodes, Code Nodes, IF routing)
· Anthropic Claude API (Sonnet 4.5) — powers all 4 AI agents via HTTP Request nodes with distinct temperature settings
· Airtable — intelligent content hub: dual-table database (Content Ideas input → LinkedIn Posts output) + mobile form for iPhone input + AI-enhanced fields for engagement rate and performance tracking
· Slack — notification delivery with formatted reports (brand voice score, revised post, hooks, timing, engagement starters)
· GitHub — version control for all system prompts, brand voice assets, and documentation (9 commits, V1→V5)
· Claude Code (VS Code extension) — primary development tool for prompt engineering, code generation, debugging, and git management
· Wispr Flow — voice-to-text dictation for mobile idea capture on iPhone
All code, prompts, and documentation were generated through natural language conversation with Claude Code — often instructing in German, receiving English code output.
METHODOLOGY
1. Skill first, then system: I started by building a Claude "LinkedIn-Post Writing-Skill" to deeply understand the content creation process. Then I decomposed that skill into 4 specialized agents — following Sara and Tyler's approach — each with XML-structured system prompts containing explicit role definitions, guardrails, and collaboration models.
2. Iterative, usage-driven development: V1 = manual trigger + 4 agents. V2 = scheduled polling + Airtable integration. V3 = improved data lifecycle. V4 = multi-date timing + exact post titles. V5 = dual-mode routing + mobile input + error resilience. No speculative features — every version responded to a real problem encountered during actual use.
3. Vibe-coding with Claude Code: Built the entire system as a non-technical founder. All JavaScript, XML-structured agent prompts, and documentation generated through natural language conversation with Claude Code — often instructing in German, receiving English code output.
4. Brand voice engineering: Created a comprehensive voice identity system (23KB XML + Voice Guide + 8 calibration posts + 5 knowledge base files + business context) that agents use to write in my authentic voice. The QA Editor acts as final guardian with rubric-based scoring and AI language red flag detection.
LEARNINGS
1. Start small, then add complexity: I started with normal chats in ChatGPT, then iterated with Claude and Perplexity, then built a Claude skill and a voice guide, then my automated workflow — always from simple to complex. Every version was fully working before I enhanced it. I already plan to use my LinkedIn posts to write a book, continuing the same iterative process.
2. What Sara and Tyler taught us was invaluable. Only once you try and apply their methodology — thinking in agents, systems, and workflows — do you get the real value. Without their classes and the foundation work with the repo and tools, I would not have built a production system in under a week.
3. Build your content brain early: After my first 5–8 LinkedIn posts, I started tracking everything in Airtable — not just text, but performance data, content pillars, hooks that worked, timing patterns. This "second brain" is now the foundation my automation draws from. The pipeline doesn't just produce content — it grows the brain that fuels future content.
4. Mobile capture changes everything: Before the Airtable shortcut on my iPhone, ideas were lost. Now I dictate in 30 seconds and the pipeline runs automatically. The best content ideas come when you're not at a desk.
5. Non-technical founders can build production systems with AI: With Claude Code as a development partner, I went from zero coding background to a fully automated multi-agent pipeline with conditional routing, error handling, and API integrations — in 5 days.
RELEVANT URLs
- GitHub repository: https://github.com/lionmic78/mindvalley-ai-mastery-students (private repository — access available on request. Contains all 4 agent system prompts, brand voice assets, Format Agent code, and full commit history)
