AI Research — Perplexity, NotebookLM, Claude long-context
AI-Powered Research
Turn any topic into deep, cited expertise in under an hour.
The 3 research superpowers
- 01Perplexity — real-time answers WITH sources. It searches and cites so you can trust it.
- 02NotebookLM — upload 50 PDFs, get a chat that only answers from YOUR docs. Ask hard questions.
- 03Claude (Sonnet 4/Opus) — 200k+ token context. Paste an entire book, ask anything.
- 04ChatGPT Deep Research — 30-min autonomous research on your behalf. Ships a full report.
Right tool, right job
Anything with 'what's the current…', 'who's the biggest…', 'compare X vs Y'. Real-time internet + citations. First stop for market research.
You have specific sources (research papers, reports, transcripts). Upload them. Ask questions ONLY grounded in those docs. Generates podcasts, briefings, Q&A. Insane for exam prep.
Guided demo: Perplexity
Pick a topic, then tick each step as you finish. We move to NotebookLM when the room hits 80% done.
Open Perplexity & sign in
Click 'Open Perplexity' (top-right of this slide) → sign in with your personal Google account. Free plan is fine.
Ask the landscape question
Paste this into Perplexity and read the answer + sources it cites.
Dig into ONE citation
Click the source that surprised you most. Skim it for 60 seconds. Copy the URL into your notes doc.
Ask the sharp follow-up
Perplexity remembers the thread. Ask this next to force it deeper.
Save 3 sources for NotebookLM
Pick the 3 best sources across your two answers. Copy their URLs into a plain doc. We'll upload them in the next demo.
What's the BEST reason Perplexity beats plain ChatGPT for market research?
Guided demo: NotebookLM
Same topic. Upload the sources you saved from Perplexity. Turn them into a private expert + a podcast.
Open NotebookLM & create a new notebook
Click 'Open NotebookLM' → sign in with the SAME Google account → hit 'Create new notebook'.
Upload your 3 sources
Paste the URLs from the Perplexity demo, or drop the PDFs you downloaded. Wait for the little green checkmarks.
Ask a question ONLY it can answer
Ask this in the chat. NotebookLM will only answer from YOUR sources — and cite the exact paragraph.
Generate the Audio Overview (podcast)
In the right panel, click 'Audio Overview → Generate'. Give it ~3 minutes. Two AI hosts will discuss YOUR sources.
Ask one 'so what' question
Type the last question. Screenshot the answer for your project kit.
You just used both. Which one do you reach for first, and why?
Which tool surprised you more, and what did it get right that you didn't expect?
Where did either tool sound confident but wrong? How would you catch it?
1 topic → 1 report in 45 minutes
Claude: 'Give me the 15 questions I MUST answer to understand [MARKET] in Pakistan.'
Ask each question in Perplexity. Save the answers + top 3 sources per question.
Drop all the source PDFs / articles you found. Now you have a private expert.
Claude: 'Here are all my notes. Write me a 5-page market brief that a VC would respect.'
The market analyst
You are the world's best market analyst. My business idea: [PITCH]. I need a report covering: (1) TAM/SAM/SOM for Pakistan, (2) top 10 competitors with pricing, (3) the 3 biggest customer pain points, (4) 5 non-obvious trends changing this market, (5) what would kill this business. For each claim, cite a source I can click.
Run in ChatGPT Deep Research OR Perplexity Pro. Come back 30 minutes later to a full report.
Turn any topic into a podcast, on demand.
Upload 5 articles about your market. Click 'Audio Overview'. In 3 minutes: two AI hosts having a genuine conversation about YOUR sources. Perfect for the walk home.
Research hygiene
- 01Trust NOTHING an AI says with a number in it — cross-check with Perplexity's sources
- 02Screenshot every important source (URLs die, papers move)
- 03'Confidently wrong' > 'obviously wrong' — the calm mistakes will bite you
- 04One prompt = one question. Don't ask 5 things at once.
Live activity: 45-min competitor report
Ship a real, cited 3-page competitor report by end of class.
- 01Pick 5 real competitors in your space (mix of global + Pakistani).
- 02Perplexity → pricing, users, launch date, funding for each.
- 03NotebookLM → upload their homepages + top 3 reviews.
- 04Claude → 'synthesize into a 3-page report + one 2×2 map.'
- 05Ship as PDF. Read it out loud. Fix anything that sounds like slop.
An agent is an AI that can DO — not just talk.
It reads, decides, and clicks on your behalf. Repeatedly.
One replies. The other acts.
You type. It answers. You do the next step.
You give a goal. It plans steps, uses tools (search, email, code, browsers), and finishes the job.
No coding required
- 01Custom GPTs — package a prompt + files + tools
- 02Zapier / Make — connect apps, trigger AI, act
- 03n8n — visual automation, self-hosted
- 04Claude Projects — context + instructions in one place
4 parts every agent needs
'When a new email arrives…' / 'Every morning at 9…'
Data it reads: the email, a spreadsheet, your calendar.
The prompt that decides what to do.
Reply, post, save, notify, buy.
The agent-designer prompt
Design a Custom GPT for me. Purpose: [WHAT IT SHOULD DO]. Target user: [ME OR CUSTOMER]. Give me: (1) name + description, (2) system instructions in full, (3) 4 starter conversation prompts, (4) knowledge files I should upload, (5) which tools to enable (web, image, code).
Actual builds by teens
- 01Homework-explainer trained on their own textbook
- 02IG DM auto-responder for a small brownie shop
- 03'Explain this scholarship' PDF-reader agent
- 04Daily news summarizer for their parents
You are responsible for what your agent does.
Set limits. Log everything. Never let it send money or messages without a check.
Bonus drill: Agent Blueprint
Design (not build) an agent that saves you 3 hours a week.
- 01List every task you repeat weekly.
- 02Circle the ones that follow rules.
- 03Pick one — design its Trigger, Context, Brain, Action.
- 04Sketch it on paper. Share with the room.
What's the smallest possible version of your product you could put in a friend's hands tonight?
Cut features until it's embarrassing.
What's the ONE thing it must do to be useful?
What can be fake / manual for now?
You're building on a no-code tool. Something breaks. First move?
You just did what junior analysts take a week to do.
Tomorrow: turn all this into a real, live product.