THE ECONOMIC TIMES

AI for Marketing Leaders

Masterclass

COMPLETE CLASS NOTES • DAY 1

A structured, screenshot-by-screenshot walkthrough

Before You Begin: How to Read These Notes

These notes cover the Day 1 session of the Economic Times AI for Marketing Leaders Masterclass, taught by Rishi Jain. The

How this document is organized

Part 1 — The Foundations: what AI actually is, and the vocabulary you need (tokens, context windows, models).

Part 2 — The Tool Landscape: every major AI tool, what each is genuinely best at, and how to verify claims yourself.

Part 3 — Talking to AI Well: prompt engineering, the CRAFT framework, and avoiding “AI slop”.

Part 4 — Building Blocks of Automation: Skills, Connectors, Plugins, and Agents explained from the ground up.

Part 5 — The Hands-On Build: step-by-step, screenshot-by-screenshot, building and deploying a real website.

Part 6 — Managing Cost & Staying Safe: token discipline and the security habits that protect you.Part 1 — The Foundations: What AI Actually Is

1.1 AI is a Prediction Engine

The single most important idea to start with: at its core, modern AI is a prediction engine. That is the whole trick.

“How are ______?” — and instantly your brain filled in “you”. Then “I’m fine.” That auto-completion in your head is exactly what an AI language model does, except it does it at massive scale.

The core mechanic

A Large Language Model (LLM) like ChatGPT or Claude predicts the next word, one piece at a time.

An image model predicts the next pixel.

A video model predicts the next frame.

Same underlying idea — prediction — just in different mediums.

Going one level deeper (so you truly understand it)

“Next-word predictor” is true, but it is the floor, not the ceiling. The genuinely interesting part is what the model had to learn in order to predict well. To reliably guess the next word across law, code, Hinglish, sarcasm, and technical docs, the model ends up building internal representations of grammar, facts, and even rough models of how the world works.

Prediction is the training goal; intelligence is the side-effect. A useful analogy: saying “a chess grandmaster just predicts the next move” is technically true — but to predict that move well, they had to internalise strategy, patterns, and endgames. The word “just” hides all the good stuff.

Why two AIs give different answers to the same question

“What is physics?” gave “nature” from one model and “science” from another. The takeaway presented was that different models predict different next words. That’s the right spirit but the real reasons they differ are worth knowing precisely:

• Different training data — each model learned from a different slice of the internet, books, and code.
• Different fine-tuning — Anthropic and OpenAI shaped their models differently after the base training.
• Sampling / temperature — a bit of built-in randomness means even the same model can answer differently twice.

One small precision for when you make content from this: the model weighs the whole sentence, not literally just the last word. “What is physics” is understood as a complete request otherwise the model wouldn’t know whether you wanted a definition, a translation, or a rhyme.

1.2 The Two Types of AI: ANI vs AGI

Type	What it means
ANI — Artificial Narrow Intelligence	Historically: AI that is very good at one specific thing (e.g. only translation, or only playing Go). This is the AI that exists and works today.
AGI — Artificial General Intelligence	Human-level ability across the board — able to reason and make its own decisions across any domain, potentially placed inside humanoids. This does not exist yet.

1.3 The Three Levels of Working With AI

A clean way to see where the whole field is heading — organised by who is driving:

1. Prompt engineering — you drive. Every step is you typing instructions to get a good output.
2. Gen-AI tools — the tool drives one task. Dedicated tools built on top of models (for writing, images, video, research). You still orchestrate the sequence.
3. Agents — the agent drives the whole workflow. It connects multiple tools and prompts, and runs autonomously: it plans, picks which tool to call, checks its own output, and retries. You just give it the goal.

The keyword that separates Level 2 from Level 3

Autonomy over the loop. An agent decides what to do next, not just how to do one thing. This is exactly the pitch behind Salesforce’s Agentforce, and it’s the thread that runs through the rest of this class.

1.4 What is an LLM? (and the two words you must master)

LLM = Large Language Model. It is trained on a huge amount of data, talks in natural language, and “understands” your queries. Think of the model as the brain. The top LLMs today: Claude, ChatGPT (GPT), Gemini, plus Grok, Llama, and DeepSeek.

1.5 Tokens — the thing to master today

This is the concept the class said to “master today”, because everything about cost and memory relies on tokens.

What actually happens under the hood

When you type “hi how are you?”, the model doesn’t see letters. It converts your text into numbers (and ultimately into 0s and 1s at the computer level). Each little piece is a token.

“hi how are you?” becomes exactly 5 tokens — each coloured piece is one token.

The model literally works with those numbers (3686, 1495, 553, 481, 30) — not words.

You write words; the model reads numbers. That gap is the whole point.

The rule of thumb: 1 token ≈ ¾ of a word

1 token ≈ three-quarters of an English word, and reversed: 1 word ≈ 1.33 tokens. Both directions describe the same fact:

• ¾ word per token — thinking from the model’s side.
• 1.33 tokens per word — the one you use to estimate cost (“how many tokens will my 1,000-word prompt use?”).

Pricing shown per million tokens — the unit “MTok” means one million tokens.

Why ¾ and not 1-to-1? The tokenizer breaks words into sub-word pieces for efficiency. “hello” is 1 token; “Salesforce” is likely 2 (Sales + force); “Agentforce” 2–3. Hinglish and Hindi words often use more tokens per word because the tokenizer was trained on mostly-English data — so Hinglish content costs slightly more at API scale.

You pay for tokens: input + output

Claude model pricing. Note Opus 4.8: $5 / MTok input vs $25 / MTok output — output is far more expensive.

Why output is ~5× more expensive than input

Reading is cheap; generating is expensive. The model finds it computationally harder to write text than to read it. “MTok” = one million tokens, so $5/MTok = $0.000005 per token — practically free for individual use. Cost only becomes real at scale: bulk agents, thousands of API calls, enterprise workloads.

1.6 The Context Window — AI’s Working Memory

Context window = the AI’s working memory. It is how much the model can “remember” within a single chat — your prompts, its replies, the whole conversation history, uploaded documents, and system instructions all count toward it.

A usage meter showing “2% used” of the session — a live view of the context window filling up.

Think of it like RAM

Powerful, but temporary. Close the chat and it’s gone (unless a memory feature is on).

Everything loads into it: your prompt, the model’s reply, history, system instructions, uploaded files.

When it fills up, the model starts “forgetting” the earliest things — that’s not a bug, it’s the design.

Seeing the limit in action

Asking for a 10,000-word essay — as it generates, the model begins forgetting its earlier words.

ChatGPT’s free version holds roughly 6,000–12,000 words of context. Ask it for a 10,000-word essay and it nearly fills the whole window in one shot — so by the time it reaches the end, the beginning is already fading out. The output drifts: tone shifts, characters change, repetition creeps in. That’s not the AI “going crazy” — its RAM is full.

Model	Approximate context window
ChatGPT Free	~6,000–12,000 words
ChatGPT Plus	~128K tokens (~96,000 words)
Claude	~200K tokens (~150,000 words)
Gemini 1.5/3 Pro	up to ~1M tokens (~750,000 words — roughly a 500-page book)

Practical rule

Don’t write a novel with free ChatGPT. It forgets the start of a long piece by the time it finishes — exactly like reading something at night and forgetting it by morning. For long projects: start fresh chats, or re-paste the key information.

Part 2 — The Tool Landscape: Who’s Best at What

2.1 Claude

• Creative writing — strong, though the class claimed ChatGPT wins for marketing (scripts, reel scripts, captions).
• Coding & AI agents — Claude understands the foundations of coding really well; excellent as a coding assistant and for agentic workflows.

2.2 ChatGPT

Pitched as the all-rounder, and specifically strong for marketing content (scripts, captions). It also has native image generation built in.

2.3 Gemini

• Deep Research — click the “Deep Research” mode. Example command from class: “a new product I can launch in India that’s viral in Korea with a huge audience base.” It doesn’t just scrape the top 10 results — it checks many sources and synthesises them.
• Images & video — via Google’s specialised models.

2.4 Perplexity

Search anything and it attaches citations to its claims. Pitched as especially useful for medical and finance — fields where you must check whether the source is trustworthy before you use the information.

2.5 Grok

Good at trending topics. Its edge comes from live X (Twitter) data — it sits inside the platform, so it catches what’s blowing up right now faster than models that need a fresh web search.

2.6 Meta AI

Important because it’s integrated into WhatsApp and Instagram. Meta’s weapon isn’t the smartest model — it’s distribution. Roughly 3 billion+ people are already inside these apps, so the AI is where people already are, with zero friction.

2.7 Higgsfield — an aggregator, not an LLM

Higgsfield is a video/image generation platform. Crucially it also acts as an aggregator — a single place to access many image and video models (Seedance, Nano Banana, etc.) so you don’t need separate accounts for each.

2.8 The whole landscape, at a glance

Tool	Genuinely best at	What it is
Claude	Coding, agents, reasoning	Anthropic — text brain (LLM)
ChatGPT	All-round, native image-gen	OpenAI — text brain (LLM)
Gemini	Deep research + Nano Banana / Veo	Google — text brain (LLM)
Perplexity	Cited answers (medical/finance)	Text brain — verify the source
Grok	X real-time trends	xAI — text brain (LLM)
Meta AI	Distribution (WhatsApp/Insta)	Meta — text brain (LLM)
Higgsfield	Video/image generation	Aggregator — NOT an LLM

2.9 LMArena — the leaderboard that settles “which is best”

This is the tool that lets you verify every “X is best” claim yourself, with data instead of opinion. LMArena (lmarena.ai) is a public leaderboard that ranks AI models by category — Chat, Code, Image, Video, Math, Creative Writing, and more — based on millions of blind human votes.

2.10 Where to find ready-made Skills (marketplaces)

Skills (covered in detail in Part 4) have grown into a whole ecosystem — like an App Store for AI workflows. Two marketplaces were shown:

skills.sh — “The Open Agent Skills Ecosystem.” Nearly 1 million skills, with a live leaderboard.

skills.sh — install any skill with one command (npx skills add <owner/repo>), works across multiple agent platforms, and ranks skills by installs. Notice that Anthropic’s official “frontend-design” skill sits near the top with 700K+ installs — the same skill available inside Claude.

Why this matters

Skill marketplaces are multiplying, and the SKILL.md format is becoming a de-facto standard across platforms. Before building a skill yourself, search these first — someone may have already solved the exact workflow you need.Part 3 — Talking to AI Well: Prompt Engineering

3.1 What prompt engineering really is

Prompt engineering = the ability to talk to AI in the smartest way possible. The insight underneath: AI doesn’t need better answers — it needs better questions. A better question produces a dramatically better answer.

3.2 The CRAFT framework

Rishi’s five-step process for crafting strong prompts:

1. Craft — write from a template.
2. Refine — look at the output and improve the prompt (and ask for 3 variations — see below).
3. Audience — tailor it for who it’s for.
4. Feedback — note what worked and what didn’t.
5. Track — iterate over time.

The prompt template

The core prompt template shown in class.

The template (copy this)

Act like a [specific role] working on [exact task or goal].

Here is only the information that truly matters:

[3 bullet points max of critical context]

Why this template works — technically

• “Act like a [role]” gives the model a persona early in the context, so every following word is predicted through that lens. The role acts as a filter on the entire output.
• “3 bullet points max” is a genius constraint. Too much context confuses the model — irrelevant tokens still carry weight. Three bullets maximise the signal-to-noise ratio. Less context, better output. Counter-intuitive but real.

Seeing it applied

The template applied live in Claude (Sonnet 5, Medium effort) — “Act like a senior meta ads engineer with 15+ years…”

Same request, completely different quality. Compare a weak prompt (“Write me a LinkedIn post about AI”) to the template applied with role, audience, tone and goal specified — the second produces something usable, the first produces generic filler.

3.3 Pro tip: always ask for 3 variations

Part of the CRAFT process — always ask the AI for 3 variations. The reason ties back to Part 1: the first output is the model’s safest, most average prediction. Variations 2 and 3 push it into “wild angles” — unexpected hooks and reframes you’d never have thought of.

The pattern

Variation 1 → expected. Variation 2 → interesting. Variation 3 → often the one. “The AI’s first answer is average, because average is what’s most predicted. Ask for the third — that’s usually the gold.”

3.4 Avoiding “AI Slop”

AI slop = low-effort, mass-produced content that sounds like AI. It’s an attitude problem, not a tool problem. The fix is to make output conversational and unmistakably yours.

Tell-tale signs of slop

• “In today’s rapidly evolving landscape…” / “It is important to note that…”
• Bold words in every sentence, bullet points for everything.
• “Certainly! I’d be happy to help…” and em-dashes everywhere.

How to fix it

• Tell it explicitly: “write like you’re texting a friend, not writing a report.”
• Give a forbidden-words list: “never use leverage, utilize, evolving landscape, it’s important to note.”
• Paste your own example: one paragraph in your style, then “match this tone exactly.”
• The read-aloud test: if it sounds awkward spoken, it’s slop.

Why this matters more than ever

AI made slop easier to produce — which made genuinely good content MORE valuable, not less. When everyone is posting 50 AI posts a day, one real, researched, conversational post stands out instantly. The value of a distinct human voice went up, not down.

3.5 Global Settings — teach the AI once

Every instruction you repeat in every prompt should go into Global Settings / Custom Instructions instead. In Claude: Settings → General → “Instructions for Claude”. These rules apply automatically to every new chat.

Claude’s global “Instructions for Claude” box — here set to “No em dashes ( no strictly ).”

What to put in global settings

“No em dashes” · “Write in Hinglish, conversational tone” · “No bullet points unless asked”

“Never use words like leverage, utilize, seamlessly” · “Always give 3 variations”

“Never start with ‘Certainly!’ or ‘Great question!'”

Craft once, use forever — this is a productivity extension of the CRAFT framework.

3.6 Privacy — turn off model training

Before using any AI tool seriously, open its privacy settings and turn off model training. In Claude: Settings → Privacy → “Help improve our AI models” → OFF.

Claude Privacy settings. “Help improve our AI models” should be turned OFF for confidential work.

Do this on every tool

ChatGPT → Settings → Data Controls → “Improve the model for everyone” → OFF

Gemini → myaccount.google.com → Data & Privacy → OFF

Especially important if you paste client data, confidential project info, or business strategies. Otherwise your strategies are training someone else’s AI — for free.

3.7 JSON prompting

Claude (and LLMs generally) parse structured formats like JSON better than plain paragraphs, because structure removes ambiguity about what is what — role, task, constraints, examples — each clearly bounded.

Plain vs JSON prompt

Weak: “Write a LinkedIn post about Agentforce for Salesforce admins, conversational, under 200 words, with a hook.”

JSON: { “role”: “Salesforce content creator”, “task”: “LinkedIn post about Agentforce”, “audience”: “Salesforce admins”, “constraints”: { “tone”: “conversational”, “length”: “under 200 words”, “structure”: “must include hook” } }

An honest nuance: JSON isn’t always best. For simple one-off requests, plain English is fine. JSON shines when the prompt is complex, has multiple constraints, or will be reused like a template.

Part 4 — The Building Blocks of Automation

This part assembles the vocabulary you need to build agents: Skills, Connectors, Plugins, and finally Agents and Agentic AI. Each builds on the last.

4.1 Skills

A Skill = a reusable prompt or SOP (standard operating procedure) that solves one specific workflow in your company. It follows your instructions to the point, the same way, every time.

The 3 parts of a Skill

1. Name — identifies the skill.
2. Description — what the AI reads to decide whether to trigger this skill. Vague description = triggers at the wrong time (or not at all).
3. Context — the actual instructions/SOP that run when the skill fires. Incomplete context = inconsistent output.

The clever way to build a skill (the 79-question trick)

ask Claude/ChatGPT to interview you with ~79 questions about how you think, speak, and operate. It took ~2 hours and produced a “brain imprint” document — the foundation for AI “thinking like you”. This is context engineering made concrete. (When you get such a document, always read it back to verify the AI captured you correctly, test a couple of real outputs against your actual voice, and refresh it every few months as your style evolves.)

4.2 Context Engineering

Context engineering = setting the full context of a workflow so the AI operates within it — your “second brain”: how you talk, how you want posts to sound, your voice on the internet. Done well, the AI starts thinking like you.

Layer	What it is	Example
Prompt engineering	One-off, per-message crafting	“Act like a [role]… 3 bullets”
Skills	Saved, reusable SOP — same workflow every time	“Generate daily post” with fixed steps
Context engineering	Ongoing identity/voice fed to AI	Your “second brain” — tone, style, decisions

4.3 Connectors (= plugins that give access)

A Connector links the AI to an external tool or service — Gmail, Google Drive, Calendar, Salesforce, Higgsfield, etc. In the agentic loop (Part 4.7), connectors are the “tools” that let the AI ACT in the real world.

Claude Connectors → Tool permissions. Each tool can be set to Always allow / Ask each time / Never.

Set permissions, don’t just connect

Always allow ✓ — routine, trusted, read-only actions.

Ask each time ✋ — anything that takes an action (sending email, deleting files).

Never ⊘ — block entirely.

Connecting isn’t enough — granular permissions stop the AI doing something you didn’t intend.

When a tool has no ready-made connector

Some tools don’t offer direct AI access. In those cases you have two options:

windsor.ai — a bridge platform that connects tools which don’t have direct AI connectors.

1. Use a bridge platform like windsor.ai — it connects to platforms that aren’t directly available and makes their data accessible to the AI.
2. Build a custom connector — if the tool has its own official MCP endpoint, connect directly.

Custom connectors and MCP endpoints

Adding a custom connector in Claude — e.g. Meta Ads via its MCP endpoint URL.

What is an MCP endpoint?

MCP = Model Context Protocol — a standard way for AIs to talk to external tools (like USB-C is a standard port for devices).

An MCP endpoint is the specific address (URL) where that connection happens, e.g. https://mcp.higgsfield.ai/mcp.

The endpoint tells the AI what it can do (its tools) and actually executes those actions against the real service.

Your Gmail connector has an MCP endpoint behind it too — when you say “check my inbox,” Claude calls that endpoint, which talks to Gmail and returns the data.

Safety — custom connectors are NOT “safe by default”

The Add-connector screen itself warns: Anthropic does not control which tools developers make available and cannot verify they’ll work as intended. So only add connectors from trusted developers, and confirm a URL genuinely belongs to that company before pasting it — especially since it may ask for OAuth credentials. Verify before you trust.

4.4 Plugins = Skills + Connectors bundled

The mind-map from class: a Plugin bundles Skills + Connectors into a single “specialist”.

A Plugin is a specialist packaged to do one job well — it bundles Connectors (access) and Skills (SOPs) together. Example: a “Creative Plugin” = Higgsfield + ElevenLabs connectors, plus JSON-prompting and static-ad-copy skills.

The whole hierarchy so far

Connector = access to one tool.

Skill = one way of doing a task.

Plugin = both bundled into one specialist — call it once, everything’s ready, nothing to connect manually.

Analogy: Skill = recipe. Connector = access to ingredients. You need both — knowing the recipe is useless if the fridge is locked.

4.5 Adding a voice: ElevenLabs connector

Add the ElevenLabs connector and the AI gains voice/audio generation — text-to-speech, voiceovers. This completes the connector-stacking pattern: Higgsfield adds visuals, ElevenLabs adds voice, so a full pipeline (script → visuals → voice) can run in one place.

4.6 Artifacts — shareable outputs

The Claude Artifacts gallery — created outputs saved, hosted, and shareable.

Claude Artifacts are created and saved on claude.ai. Publish one and it’s hosted on claude.ai with a shareable link; anyone can open and remix it. It’s like making a small interactive website without learning to code.

4.7 Agents and the Agentic Loop

The Agentic Loop: an AI that THINKS, ACTS and REPEATS until the job is done.

A normal chatbot answers once and stops. An agent keeps going until the goal is complete, running a loop:

1. THINK — reason, plan, pick a tool.
2. ACT — use a tool in the real world (read/write files, run code, search web/API).
3. OBSERVE — check what actually happened (did it work?).
4. REPEAT or DONE — if not right, loop back to THINK; if the goal is met, deliver the final answer.

It’s an AGENT when it can

Decide the next step itself · Call tools (not just chat) · Read its own results · Loop, retry & self-correct · Stop only when the goal is met.

Effort levels are the agentic loop applied to answers

Claude effort levels: Low, Medium, High (default), Extra, Max.

ChatGPT effort levels: Instant, Medium, High, Extra High, Pro.

Both platforms have an Effort setting. Higher effort = the model runs more THINK-ACT-OBSERVE cycles before answering: it thinks more, checks its own work, and retries. More effort = more agentic looping = more tokens consumed (because each loop generates new reasoning tokens) = more time and cost — but a more thorough answer.

Task type	Effort level	Why
Quick fact, simple rewrite	Low / Instant	Task is clear, little verification needed
Daily content, normal work	Medium / High	Everyday default
Complex analysis, debugging	Extra / Extra High	Multiple sources, contradictions to resolve
Hardest problem, no time pressure	Max / Pro	Maximum self-checking

Execution vs Research — the practical decision

Execution tasks (rewrite this, format this, quick fix) → Low/Medium effort. Fast, cheap.

Research tasks (compare options, deep analysis, “is this true?”) → High/Extra/Max. Slow, thorough.

Ask yourself: do I need execution or research? Wrong effort = wasted time or a weak answer.

(Precise note: high effort doesn’t literally “average 10 websites” — it reads sources, cross-checks them, flags contradictions, then synthesises. It’s cross-verification, not mathematical averaging.)

AI Agent vs Agentic AI

An SEO agent in action: THINK → ACT (real tools) → OBSERVE → REPEAT → ranked. “It ships the page. It does not hand you a checklist.”

An AI Agent is a single specialist that does one full job (e.g. an SEO agent that plans a page, publishes it, checks the rank, and fixes it). Group several specialists and you get an Agency — performance team, SEO team, creative team, intel team, data team.

AI Agent → Agency: specialists grouped by function, each doing one job.

Agentic AI is when all those agents are connected and work toward one shared business goal, with real constraints.

Agentic AI: Agents + Tools + Skills, all aimed at a goal like “generate 100 qualified leads this week” with a “₹20k/week ad budget.”

The full hierarchy — the mental model to keep

Skill → Plugin → Agent → Agency (many agents) → Agentic AI (agency + goal + budget constraint)

The punchline: with Agentic AI you don’t give the AI a task — you give it a business outcome plus a budget and a deadline, and it figures out the rest.

4.8 What can you automate with an agent?

Concrete examples given: build/maintain an AI website; write and publish blog posts; create content and publish it (generation + distribution, not just drafting). The difference from a chatbot: a chatbot hands you a draft; an agent finishes the job — live, published, done.

Part 5A — Creative Production with AI

Before the big website build, the class covered how to generate images and videos properly — and a powerful trick for reverse-engineering winning ads.

5A.1 Meta Ads Library + reverse-engineering into a JSON prompt

Meta Ads Library (facebook.com/ads/library) is an official Meta tool that publicly shows every ad anyone is running — headlines, creative, platforms, run dates.

Meta Ads Library — browsing live ads (here, the “Master AI in 30 Days” campaign with a before/after hook).

The technique

1. Find a competitor/creator running a well-performing ad in the Ads Library.
2. Copy its structure into Claude.
3. Ask Claude to turn it into a reusable JSON prompt template with swappable placeholders.

Why this is smart

Instead of writing ad copy from scratch, you reverse-engineer a proven format into a fill-in-the-blank template — same psychological structure (before/after, urgency CTA), swappable for any offer (AI course → real-estate villa). This is essentially a Skill in disguise, and it uses the JSON prompting from Part 3.

The generated JSON prompt — with a structured bottom-zone, urgency line, CTA button, and full colour system.

5A.2 Generating images with Nano Banana

Paste your JSON prompt into Nano Banana, then attach reference images (logo, your own photos) so the output is personalised. The governing principle is “garbage in = garbage out” — the higher-quality and more varied your reference images, the better the output.

Reference-image guide — Personal Brand / avatars

The six angles to capture for a personal-brand avatar.

Angle	What it captures
1. Front shot	Close-up of the face
2. Left profile	Side view, left
3. Right profile	Side view, right
4. Full body front	Shows shape and size
5. Back full profile	Clean shot from behind
6. Teeth / smiling	A laughing / smiling shot

Why multiple angles? A single front selfie doesn’t give the AI a 3-D understanding of your face and body. Multiple angles provide depth, so the generated avatar actually looks like you.

Reference-image guide — Products

For products, capture these angles

Top shot · Bottom shot · While holding it (gives scale + real-world context) · Textures (close-up detail).

“While holding” gives the AI a scale reference; “textures” tells it the material (matte vs glossy, fabric vs metal) so the result doesn’t look plastic-y or fake.

5A.3 Video models

Model	Genuinely good at	Company
Seedance	Video + your own voice	ByteDance
Omini / OmniHuman	UGC-style ad videos, human-like avatars	—
Kling	High-quality cinematic video generation	Kuaishou

Precise version — and a safety note

Seedance is a video model (not an image model), accessed through aggregators like Higgsfield.

Kling is a general cinematic video-generation model — not a “deepfake tool”. Face-swapping is a possible misuse of any face-capable tool, not Kling’s purpose.

Its “Motion Control” feature transfers a motion-reference video’s movement onto a character image — best framed as “motion transfer” / “digital avatar animation”.

Avoid using real celebrities or copyrighted movie footage as references — that raises personality-rights and copyright issues. Use your own footage or generic/stock references, and disclose AI-generated content.

Kling 3.0 Motion Control — a motion-reference video (left) + a character image (right).

5A.4 Aggregator platforms + the script→video pipeline

Aggregator websites (Higgsfield, Magnific, Weavy) give you many image + video models in one place, so you avoid separate accounts and can compare models with the same prompt.

Seedance 2.0 accessed inside Higgsfield — reference images + audio + a script.

The full Higgsfield/Seedance setup: model, 15s / 9:16 / 1080p, credits, and a history of past generations.

The pipeline

Write the script in Claude/ChatGPT → paste it into Higgsfield/Seedance → attach reference images + audio → generate. The script-writing brain and the video-rendering brain are separate tools, chained through the aggregator. “One tool doesn’t do everything — chaining the right tools in the right order is the real skill.”

Part 5B — Cowork, Claude Code & Choosing Your Workspace

5B.1 Claude Cowork — the agent for everyday work

Claude Cowork brings Claude Code’s agentic power to non-technical work. It works on a folder — you point it at a folder and describe a task in plain English, and it plans, executes, and completes it.

A verified fact worth knowing

Claude Cowork was itself built by Claude Code — the team shipped it in roughly 10–14 days, with all of the product’s code written by Claude Code. An AI coding agent built its own non-technical sibling. (Reported widely in Jan 2026; confirmed by Anthropic’s Claude Code lead.) It’s the clearest real-world proof of the agentic loop at scale.

Cowork on web vs desktop

Capability	Desktop	Web / Mobile
Local folder / file access	Yes	No
Runs when device is closed	No	Yes (cloud)
Best for	File organization, local document work	Connected-app tasks, on-the-go monitoring

Since July 7, 2026, Cowork also runs from claude.ai in the browser (Max plan first). But local folder access stays desktop-only — so for working on files on your own computer, the desktop app is best.

Live demo — Cowork organised a Downloads folder

A messy Downloads folder, organised by Cowork into clean numbered categories (plus a _FOLDER-GUIDE).

Safety note when granting folder access

Roughly 48 hours after Cowork’s launch, security researchers demonstrated a file-exfiltration proof-of-concept: a Word document with instructions hidden in invisible white text could trigger unintended actions when Cowork accessed the folder. So treat folder permissions like a new hire’s filing-cabinet access — be careful what untrusted files sit in a folder you grant.

5B.2 Claude Code — automate, deploy, ship apps

Claude Code is the terminal/IDE tool for software. Same engine as Cowork (both run on Opus 4.8), but for developers: it writes code, fixes bugs, runs tests, deploys, and ships apps. The proof it can “automate, deploy, ship”: it built Claude Cowork itself.

5B.3 Three ways to work with AI agents

There’s the “brain” (which LLM: Claude, ChatGPT, Gemini, Grok, DeepSeek) and the “hands and legs” (how you interact). Three methods:

1. Terminal method — technical, command-line. Limitation: you can’t see your files or code unless you ask. Best avoided if you’re not a developer.
2. Native websites / apps — Claude Code’s web app, ChatGPT’s Codex. More visual than terminal, but locked to one vendor.
3. IDE (Integrated Development Environment) — easiest, because you can see your files and folders. Examples: Antigravity (Google), VS Code, JetBrains.

The key advantage of an IDE — answering “can I switch models mid-task?”

In a native single-vendor tool (Claude Code), if you hit a token limit you cannot switch to another model — it’s a real limitation.

In an IDE (Antigravity, VS Code, JetBrains) you can use different LLMs — Claude, Gemini, GPT — in the same project. Hit a limit? Switch models from a dropdown, no context lost.

If you want flexibility and a backup when limits hit, the IDE route wins because you’re not vendor-locked.

Antigravity vs VS Code — which and why?

VS Code still owns ~73% of the IDE market and is famously stable. So why is Antigravity (Google’s agent-first IDE, a VS Code fork) worth trying?

Dimension	VS Code	Antigravity
Philosophy	AI added into the editor (evolutionary)	Editor built around AI from day one (agent-first)
Stability	Rock solid, just works	Snappy, but can hit “Agent terminated” errors
Ecosystem	Full Microsoft marketplace	Native AI architecture, multi-model under one roof
Best for	Production stability	Rapid dev + multi-agent orchestration

Honest verdict

Popularity isn’t the same as best-fit. Many developers in 2026 use both — VS Code for production stability, Antigravity for rapid, agent-assisted work and its key edge: multiple AI models under one roof (so you can switch when a limit hits) plus project-level agent orchestration.

5B.4 CLAUDE.md — the project’s onboarding document

Official Claude Code docs explaining CLAUDE.md and the .claude directory.

CLAUDE.md is project instructions the AI reads at the start of every session — automatically loaded into context. It holds your conventions, common commands, and architectural context, so the AI works with the same assumptions your team does — without re-explaining each time.

Tips straight from the docs

Keep it under ~200 lines — longer files still load but reduce adherence (and cost more tokens every session).

If something only matters for a specific task, put it in a Skill or a path-scoped rule so it loads only when needed.

List the commands you run most (build, test, format). Use /memory to edit CLAUDE.md from within a session.

Analogy: CLAUDE.md is a new employee’s onboarding document — company rules, tools, conventions — read on day one so you don’t have to explain them repeatedly.

Part 5C — The Hands-On Build: A Live Website, Step by Step

This is the centrepiece of Day 1: building Agent #1 — an AEO-optimised website — and deploying it live to the internet, without writing code by hand. Follow these steps exactly and you can reproduce the whole thing.

First — what is AEO? (so the goal is clear)

AEO = Answer Engine Optimization. Where classic SEO optimises to rank in Google’s blue links, AEO optimises your content to be cited by AI answer engines — ChatGPT, Perplexity, Gemini, Google AI Overviews.

SEO works at the page level (titles, keywords, backlinks); AEO works at the fact level (clear definitions, citable statistics, well-structured sections).

They aren’t rivals — AEO sits on top of SEO. The same clean, structured, authoritative content that ranks well also gets cited more often. (Verified against multiple 2026 sources.)

The Pre-Requisites — 5 Steps

The five pre-requisites, as mapped out in class.

Step 1 — Create the folder

Agents (and Cowork) only work on a folder. Create an empty folder on your Desktop or Downloads and name it (e.g. “AIAEOWebsite”). This is the workspace you’ll later give the AI access to.

Practical tip learned the hard way

Avoid spaces in folder/project names. Use AIAEOWebsite or 1st_Agent, not “ai updates”. Spaces can cause path errors (a folder named “ai updates” showed up as “ai%20updates does not exist”).

Step 2 — Open that folder into Antigravity

In Antigravity: File → Open Folder → select the folder from Step 1. Once opened, the project appears in the left sidebar.

The folder opened as a project in Antigravity’s Explorer (here shown once files exist inside it).

Step 3 — Bring the brains (install the extensions)

In Antigravity: open Extensions (Ctrl+Shift+X / Cmd+Shift+X, or View → Extensions) and install:

Searching the Extensions marketplace for “Claude Code for VS Code” (by Anthropic).

• Claude Code for VS Code — by Anthropic.
• Codex — OpenAI’s coding agent (works side-by-side in your IDE).
• Gemini Code Assist — by Google (optional third brain).

Codex — OpenAI’s coding agent — in the extensions marketplace.

Gemini Code Assist installed — the third brain.

Why install multiple brains

With Claude + Codex + Gemini all in the same IDE, you get model flexibility: if one hits a token limit, switch to another with no context loss. This is the exact advantage of the IDE route from Part 5B.

Step 4 — Teach the folder structure

A proper folder structure is a mix of MCP config, Skills, Rules, your Commands and system prompts. It’s what keeps the AI following rules instead of scattering files randomly (which wastes tokens). The trick: give the AI the official structure as a screenshot and let it build it.

The official .claude folder structure from Claude Code docs — screenshot this.

Then prompt (with the screenshot attached):

The prompt

“I want you to create an official folder structure based on the screenshot given and also create a .env file on it.”

The prompt with the structure screenshot attached.

Pick a capable model for this (e.g. Fable — “most capable for your hardest, longest-running tasks”):

Selecting the model — Fable for the hardest work.

The agent then builds the whole structure itself:

The agent creating the folders: agent-memory, agents, commands, rules, skills, workflows, plus SKILL.md, checklist.md, MEMORY.md, etc.

The completed structure in Explorer — .claude/ with all subfolders, settings.json, .env, .mcp.json, CLAUDE.md.

Why the official structure matters

Without a structure, the AI creates files randomly and burns extra tokens figuring out where things go. A proper structure is a clear map: less confusion, fewer tokens, better rule-following. Nicely structured = the AI can “talk to you” easily.

Step 5 — The .env file (for secrets)

The .env file — where secrets (passwords, API keys, access tokens) are stored, separate from code.

.env is where you store secrets — passwords, API keys, access tokens — kept separate from your code so they’re never hardcoded.

Important — .env is NOT automatically “protected”. You must secure it.

Add .env to .gitignore — otherwise it can get pushed to public GitHub, the #1 most common security leak (people accidentally publish their API keys).

Never screenshot or share the contents of a real .env file — not even for a demo. Make it a habit.

Best practice the AI itself followed here: it also created a safe .env.example (blank template) to commit, and set .claude/settings.json to deny reading .env* so agent sessions can’t pull secrets into context.

The Shortcut: a ready-made starter repo

Instead of building everything by hand, the class shared a reusable starter kit on GitHub: github.com/rrishijain/DSAEOWebsiteBuilder. Download it as a ZIP, unzip, and open in Antigravity.

The DSAEOWebsiteBuilder repo — Code → Download ZIP.

Its README explains the purpose: a reusable starter for building and deploying static, AEO-focused business websites — it onboards a client, collects the business facts that matter for search and trust, builds the site in site/, reviews it, and deploys it to Hostinger over SSH.

The starter repo opened in Antigravity — note scripts/deploy.sh (automated Hostinger deployment) and site/index.html.

Why this template is production-grade

It already includes the official folder structure, .env.example, and .gitignore (safety handled).

It includes deploy.sh — an automated deployment script for Hostinger, so you just run it.

Next steps per the README: download, add your hosting credentials, start the prompt, answer the onboarding.

Adding a design reference (Dribbble / Nike / Amazon)

Take a screenshot of any website/product page you like (Dribbble, Nike, Amazon) and add it as a reference — the same “garbage in = garbage out” principle, now for design. Established brands use proven UX patterns, so referencing them gives the AI battle-tested design language instead of guesswork.

Dribbble — searching “website” for professional design references to screenshot.

The Build Prompt

With the design reference attached, the prompt used was:

The website-generation prompt

“Can you create me a five-page website where I am selling my healthy food business with a very wonderful unique design and no prior black, dark, or purple theme colour? Come up with light and vibrant colours and generate me a full-fledged website. I have added some reference image for the website. Use this reference image to improve your designing aspects as well.”

This is CRAFT in action: role/task (5-page healthy-food site), negative constraints (no black/dark/purple), positive constraints (light, vibrant), a reference image, and an output spec (“full-fledged website”). Then: “Once the website is ready, run Local Host and share the link.”

What “Local Host” means

Local host = the site runs on a local server on your own computer for preview (a link like localhost:3000). It’s the testing/preview stage before real deployment — like reviewing a draft before final delivery.

The agent working inside the IDE — planning, writing files, and reporting back what it built and why.

Part 5D — Going Live: Deploying to Hostinger

The local website now needs hosting to become a real, public site. Options: Hostinger or GoDaddy. The class used Hostinger.

Hostinger.com/in — sign up and take your business online.

The deployment steps

1. Sign up / log in to Hostinger (or GoDaddy).
2. Open your website → Advanced → SSH Access, and click the Enable button. SSH is what securely connects Hostinger to Antigravity for file transfer.

Hostinger hPanel → Advanced → SSH Access. The SSH details (IP, Port, Username) and the Enable button live here.

1. Back in Antigravity, open the .env file and fill in: SSH_HOST, SSH_PORT, SSH_USER, SSH_PASSWORD, DOMAIN, and REMOTE_PATH.

The .env template with the exact fields to fill — shown here with safe placeholder values, not real credentials.

1. Save — and the deploy script pushes your local site up to Hostinger. Your site is now live on your domain.

The Final Result — a live, professional website

The brief was followed exactly — light and vibrant colours, no black/dark/purple, a unique 5-page design. The result: “Verdant Pantry”, a healthy packaged-snacks business site.

The finished homepage — soft sage-green + cream palette, orange CTAs, clean hero. Exactly the brief.

The products page — “Four packs that cover most of the week,” with In-stock badges and product cards.

An About/story section — “Why the recipe changes slightly through the year” — genuinely thoughtful, non-generic copy.

Part 6 — Managing Cost & Staying Safe

6.1 API vs Subscription vs MCP — how billing actually works

The class mind-map: MCP = direct access to your existing paid plans; API = programmatic calls, billed separately.

The Claude Console (API side): organization credits, monthly spend, token volume — a completely separate billing system.

	MCP (Connectors)	API
When it’s used	Real-time, you’re chatting	Automated, scheduled, background
Whose plan pays	Your existing tool’s plan (e.g. Higgsfield credits) / your Claude subscription	Anthropic’s API — token-based, separate bill
Device	App open on laptop	Runs in the cloud, no laptop needed

claude.ai subscription (Free/Pro/Max/Team) includes the chat interface, interactive Claude Code, and remote MCP connectors — billed as a fixed monthly fee.

API is a separate track — pay-per-token, for developers building apps.

Using a connector like Higgsfield inside Claude/Cowork draws on your existing subscription/tool credits — not a separate charge. Automating via the API (scripts, background agents) is billed separately per token. (Note: since June 15, 2026, Agent SDK / headless automation draws from a separate monthly credit at API rates.)

6.2 Token Management — 3 rules to not waste money

TM1 — Chat hygiene

• Long chats cost a lot — start a new chat for irrelevant / unrelated topics.
• Don’t stretch one chat forever: context-window overflow makes the model forget AND wastes tokens (remember the 10,000-word demo).

TM2 — Skills discipline

• Don’t build skills in the 100s — keep them in the 10–20 range.
• Skills are automatically invoked in every chat, so each one’s description is scanned every time (a token cost). Too many skills = heavy background overhead on every message.

TM3 — Connector hygiene

• Always switch off connectors you’re not using — an active, unused connector wastes money through background overhead.

What you learned	The discipline it creates
Context window fills up	Start fresh chats; don’t stretch one forever
Skills are scanned every message	Keep to 10–20; only the important ones
Connectors carry overhead	Disconnect the ones you don’t use

The mindset

AI isn’t free, even on a subscription. Long chats, dozens of skills, and unused connectors are all hidden costs. Using AI efficiently is itself a skill — writing a good prompt isn’t enough.

Day 1 — The Complete Picture

Everything from Day 1, in one clean hierarchy — from the smallest concept to the biggest:

The full stack you now understand

Token — text turned into numbers; the unit of cost and memory.

Context window — the AI’s working memory; fills up and forgets.

Model + Effort — which brain, and how hard it thinks (more effort = more loops = more cost).

Prompt (CRAFT) — talking to AI in the smartest way.

Skill — a reusable SOP for one workflow.

Connector (MCP) — access to an external tool.

Plugin — Skills + Connectors bundled into one specialist.

Agent — a specialist running the THINK-ACT-OBSERVE loop until the goal is met.

Agency → Agentic AI — many agents coordinating toward one business goal with a budget.

The one line that captures Day 1: This morning we learned AI “predicts the next word.” By evening we built and deployed a complete professional website — without writing a single line of code by hand. That is what AI literacy actually unlocks.