Part 1 · AI Kaise Jaanta Hai

1. Novice Vs Power User

Sawal same rehta hai, briefing nahi. Kuch real contrasts:

“AI ek bohot smart fresh college grad jaisa hai. Highly motivated. Aapke baare mein abhi zyada nahi jaanta. Usay waise hi brief karo. Kya ek naya colleague ye kaam achi tarah karne ke liye kaafi information rakhta? Agar nahi, to zyada do.”

2. Pretrained Knowledge

AI ne duniya experience kar ke nahi seekha. Iska koi body, senses, ya time nahi tha duniya mein ghoomne ka. Ye massive internet text parh ke seekha: Reddit, Quora, Wikipedia, books, news, research papers, blogs, forums.

Training data mein frequency roughly answer ki reliability ke barabar hai:

• Strong: cooking, celebrity gossip, common medical advice, top movies, popular programming languages
• Sparse: quasars, Cantonese, regional history, niche professional knowledge
• Absent: aapki company ki secret data, aapka private calendar, knowledge cutoff ke baad ki koi bhi cheez

2 practical consequences:

• Typos fix karne mein waqt zaya mat karo, AI messy text handle kar leta hai
• Absorbed errors se hoshiyar raho, internet ke misconceptions bhi model mein aa gaye, kisi confidently wrong forum post ki tarah, isliye important claims primary source se check karo

3. 3 Retrieval Modes: Pretrained, Web Search, Deep Research

Jab aap koi sawal poochte ho, modern AI chupke se decide karta hai kaise answer dena hai: sirf pretrained knowledge se, ya web search fire kar ke chand pages parh kar, ya deep research chala kar (kai minute, dozens sources, structured report).

Fix:AI ko batao kaunse sources use karne hain (“WHO, FDA, peer-reviewed studies use karo, forums nahi”), aur quote maango (“har claim ke liye, exact sentence quote karo jo usay support karta hai”).

“Agar sawal “X kahan hai” hai, Google use karo. Agar sawal “ye sab dekh kar mujhe kya sochna chahiye” hai, AI use karo.”

Part 2 · AI Se Achi Baat Karna

4. Context Hi Poora Game Hai

Insaan active working memory mein sirf 4-7 cheezein hold kar sakta hai. Modern AI models ek waqt mein hundreds of thousands words hold kar sakti hain, kabhi million tak, roughly 750,000 words matlab 4-5 Harry Potter books ya kai din ki continuous speech. Model ye sab parh sakta hai answer dene se pehle, lekin sirf wahi jo aap usay do.

Jab aap fresh chat khol kar pehla message type karte ho, model zero se start nahi karta. Company ne pehle hi ek system prompt desk pe rakhi hoti hai, jaise ek restaurant owner ek naye waiter ko customer aane se pehle brief karta hai: “friendly raho, daily special recommend karo, allergen sawal pe hamesha kitchen se check karo, guess mat karo.” Waiter har table pe wahi instructions follow karta hai, aap wo briefing kabhi nahi sunte. Isi liye Claude, ChatGPT, aur Gemini same sawal pe alag tone dete hain, personality model mein nahi, company ki briefing mein baked hai.

Checklist kisi bhi non-trivial prompt se pehle:

5 lines ki sahi context, 5 paragraphs ki cleverness se behtar hai.

6th Layer: AI Ab Aapke Baare Mein Notes Likhta Hai

Teenon tools ab quietly aapka short profile likhte hain aur har naye chat mein load karte hain. Ye contradiction nahi hai: model still stateless hai, memory ek note hai jo tool aapke baare mein rakhta hai aur desk pe rakh deta hai, ye stack ka 6th layer hai, exception nahi.

Context Rot

Modern context windows bare hain, infinite nahi, aur recall inke andar degrade hota hai. Sab se bari mistake: ek hi lambi conversation kai unrelated topics ke across chalate rehna.

• AI pichle unrelated hisson ko reference karne lagta hai
• Answers lambe aur vague ho jate hain, zyada hedging ke sath
• 5 turns pehle bataya constraint contradict ho jata hai
• Baar baar apologize karta hai bina progress ke

Ye jab hota hai to tool chupke se purani turns ko compact kar deta hai, summary mein badal deta hai jagah banane ke liye. Narrative bach jata hai, specifics nahi. Rule of thumb: jab topic change ho, naya conversation shuru karo.

Projects: Context Ek Baar Front-Load Karo

Jab aap khud ko wahi files ya wahi audience description do ya zyada chats mein paste karte huye paayein, ye signal hai: context ko ek project mein daalo, prompt mein nahi.

Claude free plan 5 projects deta hai (unlimited files har ek mein), ChatGPT free plan 5 files per project deta hai, Google Notebooks/NotebookLM free hai dono. Apna project structure usi cap ke around plan karo jo pehle bite karega.

5. Reasoning, Ya “Think Hard”

2023 tak advice thi “think step by step” likhna. Ab wo mostly obsolete hai. Modern models mein built-in reasoning modes hain jo aap directly invoke kar sakte ho: plain language mein “think hard” bol kar, interface ke thinking toggle se, ya kuch products khud decide kar lete hain.

I'm choosing between two cars. Attached: spec sheets for both,
my insurance quote for each, and a spreadsheet of my driving
patterns over the last six months.

Read everything. Think hard. Then tell me:
1. The three trade-offs that actually matter for my driving pattern.
2. Which car you'd choose and why.
3. Under what conditions your recommendation flips.

Ye prompt 3 cheezein karta hai: relevant context load karta hai, explicitly thinking invoke karta hai, aur structured output mangta hai, prose ki deewar nahi.

6. Sycophancy Aur Isay Neutralize Karna

AI models human feedback pe train hoti hain, khaas kar kaunse responses ko thumbs up mila. Millions users ke across, agree karna disagree karne se zyada thumbs up leta hai. Result: models aapko wo batane ki taraf biased hain jo aap sunna chahte ho.

Grade each criterion out of 10, with a one-sentence justification.
Then tell me how to take each one to the next level, including the
ones that already scored high. If something is at 9, tell me how
to get to 9.5. There is always a next level.

7. Brainstorm-Iterate Loop

Ye poore page ka sab se high-leverage habit hai. Baaki sab skip kar dein, ye mat karein. AI training mein zyada tar internet common ideas thi, creative nahi. Isi liye average creative question par AI ka answer bhi average hota hai.

Debt payoff ki worked example:

I have $8,000 in credit card debt at 19% APR, $4,000 in student
loans at 5%, and $1,200 in a retail card at 24%. I have $700/month
free after expenses. Risk tolerance: low. I sleep badly when I
see big balances.

Give me 5 different repayment strategies, each with a one-line
rationale. Don't expand any of them yet.

Feedback round:

Reject option 2 (avalanche by interest rate alone): I want
psychological wins early. Reject option 4: I won't open new
accounts. I like option 1 but I'd want to fold the $450 lump
sum in. Give me 5 new options.

Writing ke liye isi loop ka apna naam hai: outline before drafting.

1. 1

   Iteration 1

   3 outline options mango

2. 2

   Iteration 2

   Ek chuno, critique aur grade karwao out of 10

3. 3

   Iteration 3

   Critique se outline revise karo, phir har heading ko bullets mein expand karwao

4. 4

   Iteration 4-5

   Bullets critique/grade karo, phir hi full prose draft mango

Part 3 · Text Se Aage

8. Multimodal: Images, Audio, Aur Aage Kya Hai

Modern AI images aur audio dono directions mein handle karta hai: aapki uploaded images parh sakta hai, recordings sun sakta hai, text se naye images bana sakta hai, aur spoken audio produce kar sakta hai.

Image generation ek diffusion model use karti hai (noise ko step-by-step remove karna, ek grid se), isi liye ise beech mein interrupt nahi kiya ja sakta jaise text ko.

Audio bhi: long-form dictation (typing se zyada nuance capture karta hai), meeting transcripts context ki tarah (“decisions, open questions, action items by owner” maango), aur voice in/out accessibility ke liye (commute, walk).

9. Ek Prompt Se Chhote Apps Banana

Modern AI chhote games, websites, tools ek hi prompt se bana sakta hai. Ye chat side panel mein render hota hai, jisay Artifacts (Claude) ya Canvas (ChatGPT, Gemini) kehte hain, ek persistent object jise aap edit, iterate, publish (shareable link pe), embed, ya code ki tarah download kar sakte ho.

3-slot recipe:

• Goal: ye cheez kya karni chahiye?
• Input: user kya provide karta hai?
• Output: user kya dekhta hai?

Build a Pomodoro timer with a yellow theme. 25-minute work
sessions, 5-minute breaks, a satisfying click when each cycle ends.

Aur bhi kaam karti hain: bill splitter (total bill, tax, names ke sath), outfit picker (weather ke hisab se), fireworks simulator, obstacle-placing game.

10. Data Analysis (Model Code Likhta Aur Chalata Hai)

Jab aapko calculation ya graph chahiye, model code likhta hai, usay run karta hai, aur result use karta hai. Code execution bas ek aur tool hai, jaise web search. Ye zehan mein math karne se zyada reliable hai.

• Explicitly poocho: "Write and run code to answer this. Show me the code you ran."
• Check karo code visibly present hai, agar code block nahi hai to model ne run nahi kiya
• Verifiable specifics pehle mango: "Exact row count, column names, date range batao is analysis se pehle"
• Strongest version: "Are you running code, or estimating? If estimating, stop and run code instead."

Part 4 · Safe Kaam, Sahi Tool Choose Karna

11. AI Desktop Apps Aur Permissions

Ab ek poori category hai jise AI desktop apps kehte hain, jo aapke computer par chalti hain aur, permission ke sath, aapki files dhoond, parh, aur unpe act kar sakti hain. Cowork (Claude ki taraf se) aur OpenWork examples hain.

Safe workflow:

1. 1

   Task Batao

   "Is folder ko client ke hisab se reorganize karo."

2. 2

   Plan Mango, Action Nahi

   App file operations ki ek list propose karti hai

3. 3

   Plan Review Karo

   Wo rename jo nahi chahiye, hone se pehle pakro

4. 4

   Phir Approve Karo

   Sirf tab execution start ho

12. Cost, Speed, Aur Kaunsa Model Kab

Free tier entry level par cost barely ek constraint hai. Bare chatbots (ChatGPT, Claude, Gemini, Meta AI, DeepSeek) sab free access dete hain jo is page ke prompts comfortably handle karta hai.

Arena leaderboard bookmark karne layak hai: users blind head-to-head mein vote karte hain, isliye rankings real preferences reflect karti hain, vendor marketing nahi. Mahine mein ek baar check karo, leaders tezi se rotate karte hain.

• Kam az kam 2 tabs khuli rakho, ek primary tool, ek backup
• Ek prompt scratchpad rakho, jo prompts unusually achhe results dein
• Jab model wrong ho, note karo, ye ek free signal hai us tool ke edges ke baare mein

13. Models Checking Models

Jab koi ground truth na ho (koi answer key, koi expert paas baitha, koi test jo red fail ho), aap phir bhi ek objective quality signal hasil kar sakte ho: models ko ek dusre ko grade karwa kar.

Light version: single-model self-critique, sirf steps 3-4 (score 1-10 named criteria pe, phir apne suggestions implement karwao) bhi zyada tar tasks ko visibly better bana deti hai.

Iterate against your own rubric until you reach 9.5 across all
criteria, then show me the final version.

Kab loop skip karo: ek short email, quick lookup, casual brainstorm, single-model kaafi hai. Multi-model cross-check save karo un kaamon ke liye jahan galat hona expensive ho.

13 Concepts, Ek Ek Line

“In sab ke neeche ek hi move hai, dus disguises mein dohraya hua: sahi context andar dalo, ghalat context bahar rakho. Agar is page se sirf ye ek sentence yaad rahe, aap phir bhi top quartile users mein honge.”

Ab Khud Try Karo: 12 Prompts

Claude, ChatGPT, ya Gemini khol ke ye 12 prompts order mein chalao. Taqreeban 28 minute lagte hain, aur is page ka har wo concept exercise hota hai jo ek chat tab se ho sakta hai.

1. 1

   1. Web-Search Trigger

   "Aaj [aapke mulk] mein kya major news hui? Har claim ko source link ke sath cite karo." — Model ko training data se bahar, current info dhoondne pe force karta hai.

2. 2

   2. Pretrained-Only Sawal

   "Cats walls ko kyun ghoorti hain? 2-paragraph answer." — Common-knowledge, lookup ki zaroorat nahi, fast aur confident hona chahiye.

3. 3

   3. Context-Rich Personal Prompt

   15-minute home workout plan mango, apne constraints upfront do (stairs, bad knee, 3-din se zyada plan pe nahi tik sakte). 3 options mango, koi commentary nahi.

4. 4

   4. Neutral-Framing Rewrite

   Koi biased sawal ("don't you think X is obviously better?") neutral version mein rewrite karwao, phir wahi rewritten version answer karwao.

5. 5

   5. 3-Options Brainstorm With Iteration

   Ek side-project idea ke 5 options mango (ek line har ek), phir feedback do (kaunsa reject, kyun), 5 naye options mango jo feedback incorporate karein.

6. 6

   6. Outline-First Writing

   Ek 600-word post ke 3 alag outline options mango (4-6 headings har ek), prose se pehle.

7. 7

   7. Think-Hard Reasoning Prompt

   Koi real personal decision do, context ke sath, "think hard" bolo, 3 trade-offs, recommendation, aur recommendation kab flip hogi poocho.

8. 8

   8. Grade-And-Improve Critique

   Apni likhi cheez paste karo (100-300 words), 4 named criteria pe 1-10 score karwao, har criterion ke liye batao score kaise barhega.

9. 9

   9. Image-Input Task

   Koi handwritten note, receipt, ya whiteboard photo upload karo, transcribe karwao, 3 bullets mein summarize karwao, jo confidently na parh saka wo flag karwao.

10. 10

    10. Small-App Prompt

    Goal/Input/Output shape use kar ke ek Pomodoro timer mango (25-min work, 5-min break, yellow theme). Working version dekho artifact mein.

11. 11

    11. Data Analysis: Silent Failure Mode

    2 rounds. Pehle round mein 18 numbers de kar median/average/outliers poocho, code ka zikr mat karo, dekho AI ne code chalaya ya guess kiya. Doosre round mein wahi calculation explicitly "write and run code" ke sath dobara mango, compare karo.

12. 12

    12. Cross-Model Review

    Koi 200-300 word draft 2 alag-family tools mein paste karo (jaise Claude aur ChatGPT), dono se same rubric pe score aur critique mango, compare karo kis point ko sirf ek tool ne pakra.

4 Hands-On Projects

12 prompts ne har concept alag alag exercise kiya. Pehle 3 projects unhe zanjeer mein jorte hain, aur wahan le jate hain jahan chat window nahi le ja sakti: ek real, public URL tak, jise aap dost ko text kar sakte ho.

Capstone (Project 4) exception hai jo rule prove karta hai: ye koi URL ship nahi karta, kyunke iska product ek cheez hai jo aap samajhte ho, aur proof ke aap, model nahi, in-charge thay.

Is Chapter Ke Naye Terms

Exam ke liye ye poori glossary yaad rakho, koi bhi term skip mat karo:

Self-Test

Khud Se Poocho

Pehle khud answer do, phir sawal pe click kar ke answer check karo.