This week’s Topic Finder & Ranker classwork led me to an important realization. LLMs are excellent writers. But they shouldn’t be the ones deciding what to say. I extended the Week 4 build into an end-to-end content engine that uses an LLM as an executor, not a decision-maker. What’s different from normal GenAI usage vs this Content writing engine? When we use ChatGPT or any other GenAI tool directly, we usually ask: - “Write a post on X” - “Make it engaging” - “Add a hook” All strategic decisions are implicit and unstable. In this system, those decisions are explicit and frozen upstream. How the system works: 1️⃣ Topic Finder / Ranker: Only topics with real “why now” signals enter the system. No random prompts. 2️⃣ Idea Classification (Editor Brain) - Before writing, the system decides: - What kind of thinking this is - The intent (authority, education, engagement) - How safe or sharp the angle should be. This prevents format and intent confusion later. 3️⃣ Packaging Strategy (Creative Director Brain) Format is chosen before content: - Carousel vs text - Slide count - Visual role per slide. The LLM does not “guess” formatting. 4️⃣ Execution Engine - Only after all decisions are locked: - Post is written - Carousel blueprint is generated - PPT is auto-created via Python and Carousel prompt is generated for non-technical users. Key learning The value isn’t the writing. The value is structuring judgment around the LLM. This build helped me understand the difference between: - Prompting better vs - Designing systems that think consistently Still learning and iterating. Sharing mainly to document how my understanding of automation is evolving.

This week pushed me beyond simply following steps into actually understanding how data flows through an AI automation. Replacing JavaScript nodes with LLM chains sounded straightforward initially. In practice, I learned that LLMs silently break implicit assumptions unless context, schemas, and validation are handled explicitly. I ran into multiple dead ends like rate limits, broken loops, partial logs, and at first it felt like I was doing something wrong. Over time, I realized those failures were exposing gaps in how I was thinking about guardrails and observability. By the end, I was able to replace the JS ranker with an LLM-based ranker, add post-LLM validation, and build a more robust topic-finding workflow. Biggest takeaway: building AI workflows is less about prompts and more about protecting data flow and designing for failure.

This week was about slowing down before building. Decomposing my LinkedIn RAG project into minimal, testable nodes made me realize how much clarity comes from defining “done” first.

In Week 2 assignment i rebuild AI Automation for Digital Marketing B2B, included my website https://digitalimran.xyz

I spent this Sunday building something interesting outside of the official assignment; a Mini RAG system for LinkedIn content using n8n + Google Sheets + Gemini. The goal: 👉 Scrape high-performing posts from top AI creators 👉 Extract structural patterns automatically 👉 Generate fresh LinkedIn posts in my voice, based on proven formats Here’s how it works: 🧩 Workflow A : Pattern Builder 1. Fetches raw posts from a Google Sheet (just post URL + content pasted manually 2. Loops through each post 3. Sends the content to an LLM and returns structured data: ✔️Hook type ✔️Hook example ✔️Post type ✔️Tone style ✔️CTA style ✔️Writing tricks ✔️Key ideas 4. Saves everything into a Pattern Library Sheet This turns long unstructured posts into a database of writing patterns. 🧩 Workflow B : Post Generator 1. I enter simple inputs: ✔️Niche ✔️Target audience ✔️Post type ✔️Topic hint 2. The system: Reads multiple patterns from the library Summarizes them into one master pattern prompt Generates: ✔️ LinkedIn post ✔️ First comment ✔️ One-line summary Then emails the output to me automatically. 🧪 Status & Screenshots ✔️ The system works end-to-end ✔️ I received my own AI-generated LinkedIn post + comment by email ⚠️ Hit a Gemini API quota after processing several posts (classic 🤣) No paid tools. No external scraper APIs. Just n8n, Google Sheets, and manual LinkedIn/X copy/paste. 🧠 What I learned today: - Studying real posts beats “write a LinkedIn post like a human” prompts. - Data beats creativity. I can replicate viral positioning based on structure and data, not luck. - Workflow chaining in n8n finally clicked today. 📅 Next steps: - Add batching + retry logic to avoid API limits - Auto-scrape LinkedIn comments for CTAs and hook variants - Export pattern library to Airtable or Notion