Video summary

Day 1 | Generative AI Mastermind - 18th July'26

Main summary

Key takeaways

Educational

Main ideas / lessons conveyed

  • Generative AI basics and how it works

    • Generative AI is framed as AI that can create text, images, and videos.
    • The “AI internals” explanation is simplified into a chain:
      • Tokenization (chopping input into pieces)
      • Embeddings (turning words into coordinate-like math space; similar meanings cluster)
      • Self-attention / Transformer mechanism (focuses on the most important parts of the text)
      • Prediction (predicting the next word)
      • Response generation (building the final output)

  • “Context engineering” as the key to better outputs (more important than prompt engineering)

    • Prompting = giving instructions.
    • Context engineering = supplying the background/info that makes instructions work well (like onboarding a smart employee).
    • A mismatch in context leads to bad output even if the model is capable.

  • A structured method to write stronger prompts (for LLM apps / GPTs / Claude projects)

    • The session emphasizes a context-engineering prompt formula including:
      • Identity (role the AI should act as)
      • World / situation (audience, business, constraints, background)
      • Task (what to do)
      • Examples / what good looks like
      • Constraints / boundaries / non-negotiables

  • Model selection matters (and saves cost/tokens)

    • Use a “difficulty ladder” style:
      • fast/instant models for quick/simple tasks
      • higher-thinking models for deeper reasoning and strategy
    • Compares multiple models’ outputs to find the best one for your use case.
    • Mentions using OpenRouter and getmulti.com to test/compare outputs across many models.

  • Moving from “tools” to building “AI systems/bots/agents”

    • The progression is presented as levels:
      • Level 1: better prompting + selecting models + using tools effectively
      • Level 2: building more advanced workflows (agents, automations)
      • Level 3: building “AI employees” / automations that run tasks regularly, including voice/video/image workflows
      • Later described as continuing beyond level 3 into more advanced agent building and “VIP coding”
    • Emphasizes orchestration: combining multiple apps/tools/models so they work together like a system.

  • Tooling examples shown

    • WhisperFlow: dictation-to-clean structured text (voice → usable docs/emails/summaries)
    • Fireflies.ai: meeting transcripts + action items
    • NotebookLM (Google): chat with YouTube transcripts (summarize + quiz generation)
    • Claude Research: deeper company/research mode
    • Supergrow: LinkedIn post generation / idea assistance
    • Happstance: finding referral connections via network (for job applications)
    • There’s an AI-for-that directory: discover tools by category/rankings
    • Claude “projects” / “artifacts”: mini-app-style instruction + file-based work
    • Tools for image/video ad creation (e.g., “fort.ai” demo)
    • VAPI / 11Labs: voice agent creation + voice selection
    • Lizer AI agent builder + Google AI Studio: agent + UI integration
    • Local AI with Ollama (privacy + offline model execution)
    • Appify + MCP examples for automating content research and transformations
    • Goose: automation that can use a computer interface (screen automation)

  • Business/career framing

    • AI is positioned as transforming the job landscape: routine work shifts to AI agents.
    • The recommended response is to become the person who can apply AI to solve problems (“AI generalist”).
    • Strong emphasis on using AI to unlock opportunities: consulting, freelancing, solopreneurship, startups.

  • Privacy

    • Enterprise setups are said to include privacy options/terms; for strict needs they suggest local models.
    • Local execution is presented as the best route for maximum privacy.

Methodologies / instructions (detailed bullet lists)

1) “Five-step” intuitive model workflow (how LLM responses are generated)

  1. Tokenization
    • Split input text into smaller pieces (“tokens”).
  2. Embeddings
    • Map words/sentences into a mathematical space so similar meanings cluster together.
  3. Transformer / self-attention
    • Identify which parts of the text are most relevant to the next output.
  4. Prediction
    • Predict the next token/word step-by-step.
  5. Response generation
    • Assemble predictions into a coherent final text output.

2) Context engineering prompt formula (used for stronger outputs)

Use these sections when prompting LLMs for best results:

  • Identity
    • Specify what the AI should be (e.g., “act as a sharp chief of staff”, “analyst”, “interviewer”, “executive communication partner”).
  • World / situation context
    • Describe audience, scenario, constraints, and the environment the AI must consider.
  • Task
    • State the specific work to perform (e.g., write an email, build a report, compare strategies).
  • Examples (good vs bad)
    • Provide what great output looks like and optionally what poor output looks like.
  • Boundaries / constraints / non-negotiables
    • Rules the output must follow (tone, length, formatting, prohibited content, no fluff, must include sections/tables, etc.).

3) Prompting “upgrade” approach demonstrated in the session

  • Start with basic context-engineering structure.
  • If you want more power/structure:
    • Ask AI to transform rough notes into the full context-engineering prompt first.
  • Add feedback loops:
    • Ask AI to check its own output (“double check”, critique weaknesses, rebuild if needed).

4) Model selection framework (to reduce token waste)

  • When using ChatGPT-style models
    • Simple tasks: lighter/instant
    • Complex thinking: “thinking”/higher reasoning
  • When using Claude
    • Choose by “effort” level:
      • quick decision: smaller/lower effort models
      • deeper strategy: highest/Opus-like models
  • Use “right model for right job” to avoid:
    • wasting tokens on hard models for simple tasks

5) Build an AI mini-bot / custom workflow (GPTs / Claude Projects)

  • Use existing “GPTs” for quick micro-tools:
    • search GPTs store and try email writers, etc.
  • For a custom bot:
    • Create a custom GPT / project
    • Provide:
      • identity + objective + instructions
      • optionally upload supporting files (notes, playbooks, datasets)
  • Test with sample inputs (“topic”, “rough notes”, etc.) before relying on it.

6) “AI generalist” problem-solving loop (implicit system)

  • Define the goal/problem.
  • Decide whether AI can solve it.
  • Provide adequate context (context engineering).
  • Choose appropriate model(s) and tooling.
  • If needed:
    • orchestrate multiple tools (web research + data + writing + formatting)
  • Iterate with verification/validation (e.g., compare against your own analytics data).

7) Voice/voice-agent workflow (high level)

  • Create voice assistant/agent:
    • select voice persona
    • paste the core “mentor” instructions prompt
    • attach session transcripts/knowledge files
  • Test with spoken questions.
  • (Later demo) connect to a phone-call workflow using voice agent infrastructure.

Speakers / sources featured (explicitly mentioned)

  1. Funni Krishna (host; also referred to as “PK” outside India)
  2. Dip (Dip; associate director / leads generative AI education; AI strategy/implementation at Outskill)
  3. Webhub / Webhuff / Web of Cicinti / Web of Soci… (presented as a mentor/teacher for the second session; the name appears inconsistently in subtitles)
  4. Billy (speaker who discusses privacy + wrap-up; “Billy” appears in the subtitles)

Mentions of tool/platform sources

  • Anthropic (Claude)
  • OpenAI (ChatGPT)
  • Google (Gemini, NotebookLM, Google AI Studio)
  • xAI / Grok (referenced as “Grock”)
  • Meta (referenced)
  • Microsoft Copilot (referenced)
  • OpenRouter (openrouter.ai)
  • getmulti.com
  • Ollama (local models)
  • Vapi.ai
  • 11Labs
  • Lizer AI
  • Happy Scribe / video-to-transcript tools (mentioned)
  • Zerodha (mentioned in the stock demo)
  • Zomato (mentioned in food ordering demo)
  • Goose (computer automation demo)
  • Appify (MCP scraping/automation demo)
  • Claude Research
  • Fireflies.ai
  • WhisperFlow
  • Supergrow
  • Happstance
  • Super prompts (prompt organization tool)
  • Super prompts / “super prompts” and “Claude artifacts/projects” (internal Outskill workflow tools)

Other references

  • Jay Gupta (referenced as an author of an article mentioned during the talk)
  • Sam Altman (referenced in a discussion about “single founder” startups)

Note: Some names—especially “Webhub/Webhuff”—appear with multiple spellings in the auto-generated subtitles.

Original video