AI Automation Mastery II: an AI LinkedIn Pipeline with Human-in-the-Loop Approval and Nano Banana Images, Apify Lead Scraping to Personalised Cold Drafts, an AI Job-Match Scorer, Telegram Bots, and the Six-Level AI Generalist Roadmap

PROCEDURE: Sheet columns post idea / instruction / generate image (y/n) / status / post ID. Schedule Trigger -> Google Sheets Get Rows (status=ready, first match) -> Google Gemini 'Message a Model' (Gemini 2.5 Pro; API key from aistudio.google.com/app/apikey) with the ghostwriter system prompt and a user prompt of idea + instruction -> IF generate image = yes -> Gemini 'Generate an Image' (Nano Banana Pro) -> Google Drive Upload File to a public folder (binary field 'data') -> Gmail 'Human Review' (approve/reject) with text and the Drive link; the run blocks until clicked -> IF approved -> IF image -> Drive Download File -> LinkedIn Create a Post (media image or none; Set up manually, Organization Support off, Person) -> Sheets Update Row status=posted + post ID; rejected -> status=canceled. Tested live twice: one rejected, one posted. Q&A upgrade: a custom Form node instead of approve/decline lets the reviewer ask for changes, and a second LLM call revises - JSON shared.

A bare user prompt returns 'I'm a large language model'; a system prompt ('You are my personal assistant Jarvis; add a motivational quote to every reply') changes persona and behaviour - the rulebook versus the message. Every AI node in the session splits the two: ghostwriter, image prompt, cold-email writer, job matcher. Pin Data freezes a node's last output so the canvas can be edited without re-running slow or paid Apify and Gemini calls.

Apify: 'actors' are scrapers in a store of ~28,000 (1,500 for Instagram alone); $5 free credit; Harvest API actors rated 4.9 with tens of thousands of users; test with 3-5 items - 500 profiles takes an hour. PROCEDURE: Manual Trigger -> Apify 'Run an Actor and Get Dataset' (LinkedIn profile search, JSON input copied from the actor's Input tab: CTOs, San Francisco + New York, 10+ years, limit 3-5) -> second Apify actor (profile scraper + email) fed the profile URL -> Google Sheets Append to 'CTO Leads' (first, last, URL, email) -> Gemini 'Message a Model' in JSON output mode {subject, body} with a cold-email system prompt for 'ClearVoice' and the lead's headline and About as context; switched to OpenAI nano mid-session when Gemini crawled -> Gmail Create Draft (subject, HTML body, recipient). Live failures: free-plan run limit ('limited to 10 runs'), re-runs multiplying items, a five-minute break to fix billing.

Schedule Trigger (5 PM) -> Google Drive Download File (resume PDF) -> Extract from File (text) -> job criteria from a Sheet (title, location, easy-apply) -> Apify 'Advanced LinkedIn Job Scraper' (Harvest API, JSON input) -> Loop Over Items -> AI agent (system prompt: analyse resume and job, give a match score, write a two-paragraph cover letter without name/address, escape special characters, JSON only) -> Edit Fields splits {score, cover_letter} -> Sheets Update Row -> IF/Switch score > 50 -> Telegram message with the link. No official Easy Apply automation exists. Shown from the whiteboard with partial demo for time.

Telegram: open BotFather -> /start -> New Bot -> display name -> username ending _bot -> copy the token -> n8n Telegram credential -> 'On Message' trigger or send action. WhatsApp needs Meta's Business Cloud API, client ID and secret, ~an hour of setup - deferred to a slide deck. JSON: workflow menu -> Download; Import from file on another account; credentials never transfer, every node reconnects; templates import with 'Use for free'. All three workflow JSONs plus the revise-loop variant were shared. Also: scraping legitimacy, Instagram/Facebook publishing via the same pattern, Adobe Express not integrable, Claude can draft a workflow JSON from a prompt screenshot.

Phase 1 builds deep skills, Phase 2 builds AI-first distribution. Levels: 1) GenAI fundamentals and prompting; 2) MCP, RAG, local LLMs; 3) image/video generation and cloning; 4) automations, voice and AI agents ('scratched the surface' today); 5) building products by vibe coding and natural language; 6) solving real problems and deploying MVPs. Framed with '80% of work taken over by agents by 2027' and used as the on-ramp to the paid six-month AI Generalist Fellowship (as-heard: $5,995 / Rs 1,95,000, scholarship to ~Rs 1,75,000, 10-15 interview slots, next batch 27 June 2026, weekend evenings IST).