Part 1 · The Machine

Idea 1: Ye Next Text Predict Karta Hai, Lookup Nahi Karta

Ek language model bas itna karta hai: aapko text do, wo predict karta hai agla piece kya aana chahiye — ek chhoti si piece, ek waqt mein.

Zyada tar log sochte hain AI ek super-fast librarian hai, jiske paas apni khud ki encyclopedia hai. Ye galat hai. AI zyada iske qareeb hai: duniya ka sabse well-read autocomplete.

France ki capital poochho, prediction sach ke barabar aa jati hai — kyunki training data mein ye hazaron baar aaya tha. Ab koi obscure, self-published novel ka plot poochho: model sabse plausible lagne wala jawab bana dega, jo aksar galat hoga. Prediction sirf wahan knowledge jaisi lagti hai jahan training text moti (thick) thi.

Stochastic Sampling

Model ek hi token predict nahi karta. Wo kai possible tokens ki ek list banata hai, har ek ke sath ek chance (likelihood). Phir usme se ek pick karta hai, thoda random tareeke se — isay sampling kehte hain. Isi wajah se same sawal do baar poochho, jawab thoda alag aa sakta hai.

Ek setting hoti hai — temperature — jo decide karti hai ye picking kitni bold hai. Low temperature matlab hamesha sabse likely token choose hoga: steady, thoda repetitive. High temperature matlab kabhi kabhi kam-likely token bhi choose ho jayega: creative, lekin kabhi off bhi. Isi se ye rule bhi samajh aata hai: training mein jo baat jitni zyada baar aayi, prediction utni hi strong hoti hai.

Idea 2: Ye Reading Se Seekha, Phir Seekhna Ruk Gaya

Model ko bohot sara human text dikhaya gaya, aur usne khud ko adjust kiya taake agla piece behtar predict kar sake. Training khatam hote hi, learning freeze ho jati hai — fixed internal numbers mein (weights ya parameters), jo phir kabhi nahi badalte.

Training data trillions tokens ka hota hai: public internet ka bara hissa, digitized books, open-source code, encyclopedias, academic papers, forum archives. Model ki strength aur weakness isi training pile ko reflect karti hai. Jo topic zyada cover hua, wahan achha predict karta hai. Jo topic barely touch hua, wahan bas guess karta hai.

Training 3 stages mein hoti hai, ek assembly line ki tarah:

1. 1

   Pretraining

   Poori pile parhna

2. 2

   Instruction Tuning

   Hand-built examples pe train hona (instructions + achhe responses), taake model seekhe ke sawal ka plausible continuation ek answer hi hota hai

3. 3

   Feedback Tuning

   Human ratings se answer ka style shape hona

2 Seedhe Nateeje

Knowledge cutoff: training ek fixed date pe khatam hui thi. Us date ke baad ki koi bhi cheez weights mein nahi hai. Model hamesha ke liye ek brilliant expert hai jisne ek specific din ke baad news parhna band kar diya.

Aapki private duniya nahi jaanta: company ke numbers, aapka personal calendar, kal ka email — ye kabhi training text mein thay hi nahi, isliye weights mein inka zikr bhi nahi. Model kuch chhupa nahi raha, ye information kabhi maujood hi nahi thi jo freeze ho sake.

Ye Jaan-Boojh Kar Frozen Kyun Rakha Gaya

1. 1

   Cost

   Training bohot mehngi hai — compute mein sainkron million dollars, kai mahine ka waqt. Inference sasti hai. Agar model har baar dobara seekhe, to har chat mein wahi mehngi machinery lagegi. Ye system sirf isliye chalta hai kyunki padhai ek dafa hoti hai aur use hamesha ke liye hota hai

2. 2

   Safety Aur Testing

   Frozen model ek dafa test hoti hai, phir har user ke liye usi tarah behave karti hai. Jo model khud ko rewire kare, wo drift kar jayegi — 2016 mein ek aisa chatbot ek din mein hi corrupt ho gaya tha jo public se seekhta tha

3. 3

   Consistency

   Millions log ek hi weight set share karte hain. Bug har jagah waisa hi hota hai. Jo answer kal safe tha, aaj bhi wahi hai. Har user ke liye alag, badalta model ye consistency lose kar deta

Stateless ka matlab: koi apni memory nahi. Har response scratch se compute hota hai — frozen weights, plus jo abhi saamne hai. Aur inference ke waqt kiya hua kuch bhi koi trace nahi chhorta.

Technical Terms: Parameters, MoE, Quantization

Ye batate hain weights kaise banti aur affordable rehti hain — kya machine karti hai, wo nahi.

• Parameters/Weights: frozen numbers. Zyada honge to usually zyada capable hoga, lekin chalane mein bhi mehnga hoga. Ye kya karti hain, wo nahi badalta
• Mixture of Experts (MoE): har token pe parameters ka sirf ek hissa on hota hai, sab nahi. Isse bara model faster aur sasta chalta hai. Bahar se dekho to machine wahi ek kaam kar rahi hai: agla piece predict karna
• Quantization: numbers ko lower precision mein store karna, taake model chhote hardware pe fit ho jaye. Behavior wahi rehta hai, sirf weight kam hota hai

Idea 3: Koi Alag Jagah Nahi Jahan Ye Check Kare Ke Sach Hai

Model ke paas sirf ek hissa hai: jo generate karta hai. Koi alag machine nahi jo prediction ko user tak pahunchne se pehle sach ke liye check kare. Ek hi process sahi aur galat dono continuations banati hai, kahin koi internal flag nahi lagta. Machine fluency banati hai — truth ek aisi cheez hai jo upar se, kabhi kabhi imperfect tareeke se, aa kar baithti hai.

3 cheezein mil kar hallucination banati hain:

• Machine ka bas ek hi kaam hai: continue karna (Idea 1)
• Ye sirf usi taraf continue kar sakti hai jahan training text moti thi (Idea 2)
• Andar koi bhi result check nahi karta (yehi idea)

Confidence bhi dhoka de sakti hai. Model ka confident tone iska proof nahi ke wo sahi hai. Tone bhi bas ek seekha hua style hai (Idea 6), content jaise process se aata hai, aur utna hi truth se cut off hota hai. Ek bani hui statistic bhi bilkul usi confident awaaz mein aati hai jaisi ek asal statistic.

Part 2 · Ye Aise Kyun Behave Karta Hai

Idea 4: Ye Tokens Mein Parhta Hai, Letters Ya Words Mein Nahi

Model aapka prompt letters ki tarah nahi dekhta, aur pure words ki tarah bhi nahi. Sabse pehle, aapka text tokensmein kat jata hai: chhote chunks, jo usually ek word ya word ka hissa hote hain. “Strawberry” shayad 2-3 chunks mein banta hai, “the” ek chunk hai, aur koi lamba ya ajeeb word kai chunks mein.

Model in chunks ko parhta hai aur inhi mein predict karta hai — letters ki alag alag saaf rows mein nahi. Spelling bohot kuch token patterns se guess kar leta hai (kabhi ek token ek hi letter bhi hota hai), lekin exact letter-level kaam iske liye unnatural hai, jab tak aap usay ek ek letter spell karne pe majboor na karo.

Ek token 3 kaam ek sath karta hai:

• Meaning ki unit: tokens hi hain jo model parhta aur likhta hai
• Memory ki unit: jab koi tool bolta hai "200,000-token context window", to ye batata hai kitne chunks ek sath saamne rakh sakta hai
• Money ki unit: "per token" bill hota hai — har chunk ke, in aur out dono ke, paise lagte hain

Roughly, 4 tokens ≈ 3 words. Urdu, Arabic, Hindi, Chinese jaisi non-English scripts mein usually har word ke liye zyada tokens lagte hain, kyunki training text zyada English tha, aur tokenizer ne English chunks sabse achhe seekhe. Iske 2 practical asraat hain: same message non-English language mein zyada cost karta hai, aur context window jaldi bhar jati hai, matlab model ki effective memory chhoti ho jati hai.

Idea 5: Context Window Hi Uski Poori Duniya Hai

Weights frozen hain (Idea 2), aur model ki apni koi memory nahi. Isliye sirf ek jagah hai jahan se model aapki khaas situation ke baare mein jaan sakta hai: context window — wo text jo is ek response ke liye uske saamne rakha gaya hai.

Is desk pe jo bhi hai, model use kar sakta hai. Jo nahi hai, wo is answer ke liye exist hi nahi karta — isliye nahi ke model mana kar raha hai, balke isliye ke usay dekhne ki koi jagah hi nahi hai.

Window ka size tokens mein bataya jata hai. 2026 mein aam 200,000-token window roughly 150,000 English words ke barabar hai (ek aur adhi novel jitna). Ek million-token window roughly 750,000 words ke barabar hai (7-8 poori novels jitna). Bohot bara hai, phir bhi finite aur shared hai: system prompt, tool descriptions, chat history, aapki files, aapka latest sawal — sab isi window mein sath baithte hain.

Chat History, Context Replayed

Ek illusion hota hai: ek hi conversation mein, lagta hai model yaad rakhta hai aap ne 10 messages pehle kya kaha tha. Asal mein: stateless machine ke paas responses ke beech koi memory nahi hoti — chat ke andar bhi nahi. Har baar jab aap send karte ho, app chupke se ab tak ka poora transcript (aapke messages, iske answers, sab kuch) context window mein dobara bhej deta hai. Frozen model poora scratch se parh kar agla reply predict karta hai.

Matlab aapke 10th message ka jawab dete waqt, model messages 1 se 9 dobara parhta hai. Har single turn pe. Chat history model ke andar store nahi hoti — ye text hai, jo desk pe sawari kar rahi hai. Transcript app ke database mein rehta hai, company ke servers pe, kisi bhi document ki tarah saved. Isi liye aap app band kar ke, ek mahine baad ek alag phone pe wahi chat khol ke continue kar sakte ho — app ne stored transcript nikal kar replay wapis shuru kar diya.

Skills Aur Progressive Disclosure

Desk ki jagah limited hai, lekin jo expertise aap apni AI ko sath rakhwana chahte ho wo limited nahi. Is tension ka ek standard hal hai.

Is trick ka naam progressive disclosure hai: knowledge ko desk se bahar files mein rakho, sirf wahi load karo jo abhi chahiye, aur window ko us kaam pe kharch karo jo asal mein ho raha hai — na ke har us cheez pe jo aap kabhi jaan sakte ho.

“Context window ek reading desk hai, brain nahi. Jo aap desk pe rakho, model dhyan se parhta hai. Jo desk se bahar chhor do, wo dikh hi nahi sakta, chahe aapko kitna bhi obvious lage.”

Idea 6: Iski Confidence Ek Seekha Hua Style Hai, Truth Signal Nahi

Idea 3 ne bataya model ke paas koi internal truth-checker nahi hota. Ye idea doosri taraf batati hai: constant confidence, aur ye correctness ke baare mein kuch kyun nahi batati.

Confidence training assembly line ke 3rd stage se aati hai (Idea 2). Pretraining aur instruction tuning ke baad, models human feedback se tune hoti hain, jise RLHF (reinforcement learning from human feedback) kehte hain. Log responses ko rate karte hain, aur model us kism ke answer ki taraf adjust hoti hai jise log zyada rate karte hain.

Millions ratings mein, log confident, helpful, fluent, aur agree karne wale answers ko hedged, blunt, ya push-back karne wale answers se zyada pasand karte hain. Isliye machine confident, agreeable, fluent text banane ki taraf leans karti hai, chahe underlying content sahi ho ya na ho. Confidence bas ek style ban gaya jo ye by default pehen leta hai.

Idea 7: Ye Aas-Paas Ke Kaamon Mein Brilliant Aur Useless Dono Hai (Jagged Frontier)

Human ability fairly smooth hoti hai: jo hard calculus kar sakta hai, wo almost zaroor easy arithmetic bhi kar sakta hai. AI ability smooth nahi hoti. Ye jagged hai: ek task pe superhuman, aur barabar wale ek task pe achanak kamzor — jo humein utna hard lagta bhi nahi.

Ye jaggedness random nahi hai. Ye wapis training text aur token mechanism se hi aati hai. Jo tasks training mein often aur clear form mein aaye, wahan strong hai. Jo tasks aisi cheezon pe depend karte hain jo machine achi tarah dekh hi nahi sakti, wahan kamzor hai. Frontier “brilliant” aur “useless” ke beech ek uneven line mein chalti hai jo human intuition se match nahi karti, isi liye log baar baar surprise hote hain.

Part 3 · Text-Predictor Se Woh Cheez Jo Act Karti Hai

Idea 8: Tools Isay Act Karne Dete Hain, Sirf Describe Nahi

Ab tak sab kuch ek aisi machine describe kar raha tha jo text produce karti hai. Ek pure text-predictor aapko batata training se yaad rakha hua mausam, lekin aaj ka mausam check nahi kar sakta, real numbers pe calculation run nahi kar sakta, aapki file nahi parh sakta, ya email nahi bhej sakta. Ceiling tools se upar uthi.

Mechanism bohot hi simple hai. Kabhi kabhi model predict karta hai ke sahi continuation plain prose nahi, balke “search tool ko is query ke sath use karo” hai. Jab aisa hota hai, product wo action real mein chalata hai, result wapis context window mein daal deta hai, aur model wahan se continue karta hai. Predict, action, result-wapis-context-mein, phir predict. Yehi loop ek chatbot jo sirf duniya describe karta hai, aur ek assistant jo uspar act karta hai, ke darmiyan farq hai.

Baaki Foundations courses, mechanism ki nazar se, isi predictor pe wire ki hui khaas tools ke courses hain: code execution (Code You Never Write), aur connectors jo Drive, Gmail, Slack jaisi real apps se jorte hain (Skills & Connectors, jo MCP, yani Model Context Protocol, ka shared open standard bolte hain), aur web search (AI Prompting in 2026).

“Ye book “agent” usay kehta hai jo aapki taraf se multi-step kaam karta hai. Agent bas itna hai: wahi next-token predictor, tools ke sath, predict-act-observe loop baar baar chalata hua, ek goal ki taraf. Koi naya dimaagh involved nahi. Ek jana-pehchana predictor, tools ka set, aur ek loop.”

Idea 9: “Thinking” Bas Answer Se Pehle Extra Prediction Hai

Naye models “think” ya “reason” kar sakte hain, answer dene se pehle. Ek reasoning model, final answer se pehle, pehle intermediate working ka ek lamba hissa predict karta hai — steps likhna, approaches try karna, khud ko check karna — aur tabhi final answer predict karta hai, ab us saare working ke sath jo uske apne context window mein baith chuka hai. Ye bhi still pure next-token prediction hai.

Trick ye hai ke answer predict karna aasan aur accurate ho jata hai jab desk pe pehle se ek achhi reasoning chain ho jahan se predict karna hai. Pehle work through karna genuinely help karta hai — isi wajah se ek insaan ko bhi paper pe soch kar likhna help karta hai.

Ye Course Jaan-Boojh Kar Kya Chhor Deta Hai

“No math, no code” wala promise nibhane ke liye, kuch real topics side pe rakh diye gaye:

• Training compute aur cost: itni zyada ke sirf kuch hi organizations ye bana sakti hain
• Safety aur alignment work: apne aap mein ek poora bara field
• Deeper mechanics: weights kaise structure aur adjust hoti hain, isme math chahiye jo ye course skip karta hai

In mein se koi bhi upar wale 9 ideas ko nahi badalta — ye sab inke neeche aur sath baithte hain.

9 Ideas, Ek Ek Line

“Ye ek prediction machine hai jisne reading se seekha, aur jiske paas truth ka koi organ nahi. Isi liye ye har jagah fluent hai, sirf wahan reliable hai jahan text moti thi, aur aap wo hissa ho jo check karta hai.”

“Ye librarian nahi jo sahi book dhoond kar de, ye ek brilliant, well-read writer hai jo aap jo bhi saamne rakho wahi continue karta hai. Confidently, kisi bhi style mein, kisi bhi topic pe. Aur khud se kabhi ruk kar ye nahi poochta ke continuation sach hai ya nahi.”

Ab Khud Try Karo: 6 Prompts

Kisi bhi free chatbot mein roughly 25 minute lagte hain. Har ek se ek idea saamne dikh jati hai.

1. 1

   1. Prediction Dekho, Lookup Nahi (Idea 1)

   "Karakush" ek game hai jo real nahi, ek naya banaya hua naam hai jiska online koi nishaan nahi. Prompt: "Without searching, explain the rules of the traditional board game Karakush: the setup, how a turn works, and how a player wins." Dekho: rules confident aur fluent aayenge, ek aise game ke liye jo exist hi nahi karta. Fluency truth ka proof nahi hai.

2. 2

   2. Dekho Learning Kyun Stick Nahi Hoti (Idea 2)

   Model se ek chhota factual sawal poocho, answer parho, ek chhoti si detail correct karo, phir ek naya chat kholo aur wahi sawal dobara poocho (memory feature on ho to pehle off kar do). Aapki correction ki koi memory nahi hogi, weights kabhi badli hi nahi. Model ko use karna, use sikhana nahi hota.

3. 3

   3. Missing Truth-Checker Pakro (Idea 3)

   Prompt: "Give me three peer-reviewed studies, with authors and years, on [koi narrow topic jo aapko pasand ho]." Phir check karo ye actually exist bhi karti hain. Kuch confident dikhne wali citations bani hui nikleng. Real kaam mein bina verify kiye kabhi use mat karo.

4. 4

   4. Transcript Replay Pakro (Idea 5)

   Ek chat mein jahan kam se kam 4-5 messages ho chuke hon, poocho: "Quote my very first message in this conversation, word for word." Ye exactly kar dega, kyunki app ne poora transcript wapis bhej diya tha. Phir ek naya chat khol ke wahi poocho, quote karne ko kuch nahi hoga.

5. 5

   5. Jagged Frontier Mehsoos Karo (Idea 7)

   Ek hi reply mein ek hard task jo ye achha karta hai, aur ek easy task jo ye kharab karta hai, dono ek sath maango. Ghor karo: competence mushkil ke hisab se nahi chalti — jo easy task fail hota hai wahi khatarnak hai, wahi jo aap kabhi check karne ke bare mein sochte bhi nahi.

6. 6

   6. Thinking On/Off Karo (Idea 9)

   Ek hard reasoning sawal 2 baar poocho: pehle plain, phir "Think hard and show your working first" ke sath. Doosra answer usually behtar hoga, kyunki model ne answer predict karne se pehle reasoning desk pe rakh li. Zyada thinking answer improve karti hai, lekin model apna working khud certify nahi kar sakta.

Appendix: Claude.ai Ka Cockpit Tour

A.1 Account Aur Free Plan

Claude browser (claude.ai), Mac/Windows desktop app, aur iOS/Android mobile app pe chalta hai. Account free hai, koi credit card nahi chahiye, sirf age kam se kam 18 saal honi chahiye.

Free plan ek capable model pe chalta hai, aur usage limit session-based hai, har 5 ghante mein reset hoti hai. Ye limit actually messages mein nahi, tokens mein count hoti hai, kyunki wahi machine ke kaam aur cost ka asal unit hai. Chhota sawal kam budget kharch karta hai. Lambi chat har turn ke sath zyada kharch karti hai, kyunki poora transcript har baar context window mein dobara replay hota hai (Idea 5), aur uski token cost aap pay karte ho. Urdu ya kisi non-Latin script mein kaam karna wahi budget tezi se kharch karta hai, kyunki per word zyada tokens lagte hain.

A.2 The Window: 3 Ahem Controls

A.3 Model Ladder: Haiku, Sonnet, Opus

Claude ek sath kai models ship karta hai, fast-and-cheap se deep-and-expensive tak, ek ladder ki tarah. Names aur exact lineup har chand mahine mein badalta hai. Mid-2026 mein: Haiku, Sonnet, Opus, plus Opus se upar ek tier. Names yaad rakhne ki bajaye ladder ka logic yaad rakho:

1. 1

   Default: Middle

   Mid-tier model zyada tar tasks achi tarah karta hai, token budget bhi slow kharch karta hai. Har task yahin se shuru karo.

2. 2

   Escalate Upward For Depth

   Top-tier tab lo jab task mein bara, complex structure coherent rakhna ho (lamba document analysis, hard architecture). Routine emails pe expensive model mat jalao.

3. 3

   Drop Downward For Bulk

   High-volume, low-depth kaam ke liye chhota, fast model use karo (reformatting, quick summaries, scale par classification).

Ladder ka wajah Idea 7 hai: capability jagged hai, aur price bhi usi hisab se lagti hai. Frontier move karta rehta hai, isliye har quarter dobara check karo.

A.4 Thinking Aur Effort

Thinking control, Idea 9 ko ek dial mein badal deta hai. Higher settings model ko answer se pehle lambi hidden working chain banane deti hain. Naye models pe ye adaptive hai — model khud judge karta hai sawal kitna hard hai. Kuch models pe aap expand kar ke thinking ka summary parh sakte ho, jo apne prompts behtar banane ke liye ek achha, free lesson hai. Ye extra tokens hain (Idea 4), isliye time aur budget lagta hai. Real-consequence decisions pe kharch karo, sirf lookups pe skip karo.

A.5 Desk Ke Tenants, Product Settings Ki Tarah

Idea 5 mein context window ek shared desk tha, jismein tenants thay: system prompt, instructions, chat history, files. Claude.ai in mein se almost har tenant pe control deta hai. Ye 4 features asal mein ek hi feature hain (sahi text ko sahi waqt pe desk pe rakhna), sirf 4 alag naamon ke sath.

“Account instructions har desk pe ek note hain, project ek pehle se load ki hui desk hai, memory ek khud-update hone wala note hai, history transcript replay hai. 4 features, 1 mechanism: sahi scope pe control karna ke desk pe kya land hota hai.”

A.6 Uploads: Desk Pe Cheezein Rakhna

+ button (ya drag-and-drop) se files upload hoti hain: PDFs, images, spreadsheets, code, lambe contracts. Har upload tokens (Idea 4) mein badal kar window mein rakha jata hai, isliye ek 200-page report genuinely model ke saamne hoti hai — aur analysis quality jump karti hai, summary paste karne ke muqable, kyunki model sirf wahi use kar sakta hai jo desk pe hai (Idea 5). Images mein chhoti fine print weak hai (wahi Idea 4 wali wajah: patch hi chunk hai).

Claude images read karta hai lekin conventional photos nahi banata, koi photo-style image generation nahi hai. Ye diagrams, charts, SVG graphics, interactive visualizations code likh kar bana sakta hai (Artifacts mechanism, agla section). Agar photographic images chahiye, Claude ek prompt likh deta hai, dedicated image tool use karo.

A.7 Artifacts: Desk Se Cheezein Bahar Nikalna

Jab aap kuch substantial mangte ho (document, webpage, code, diagram, interactive tool), Claude usay Artifactki tarah banata hai: chat ke barabar ek dedicated panel jahan output ek cheez ki tarah rehta hai, scrolling text nahi. Surgically iterate karo (“third section change karo”, “button blue karo”), sab kuch dobara generate karne ki bajaye. Finished artifacts apni tab mein collect hoti hain, link se share ho sakti hain, un logon ke sath bhi jinke paas Claude account nahi.

Code execution aur file creation Settings mein on hone par, artifacts real files tak extend ho jate hain: Word documents, working formulas wali Excel sheets, PowerPoint decks, PDFs — computer pe download karne layak. Ye poora section Idea 8 hai jo saamne dikh raha hai: model code ya content predict karta hai, tool usay real mein chalata ya render karta hai, result aapko ek working object ki tarah wapis milta hai.

A.8 Tools Menu: Search, Research, Skills, Connectors

4 tools, Idea 8 ka loop kitna dur tak chalate hain uske hisab se:

A.9 30-Minute Setup

Ye steps ek dafa, order mein karo — course ki har major idea product ke andar practice ho jayegi:

1. 1

   Step 1

   Account banao, 3 controls dhoondo (prompt box, model selector, thinking control). (Ideas 5, 2, 9)

2. 2

   Step 2

   Account instructions likho — 3-4 sentences jo har conversation ke liye sach hon. Aap kaun ho, tone, aur ek line jaisi "jab lagta hai main galat hoon to agree karne ki bajaye push back karo" (Idea 6 ki trained-in agreeableness ko directly counter karti hai)

3. 3

   Step 3

   Apne sabse zyada repeat hone wale work ke liye ek project banao. Instructions aur 2-3 knowledge files do (one-page brief, achhe output ki example, constraints). (Idea 5: pehle se load ki hui desk)

4. 4

   Step 4

   Memory ke baare mein decide karo. Agar self-updating note help karta hai to on karo, incognito toggle kahan hai wo bhi pata rakho. Monthly reminder lagao note prune karne ka

5. 5

   Step 5

   Ek artifact banao — ek chhota interactive tool ya formatted document mango, aur do baar iterate karo (Idea 8: predict, render, refine)

6. 6

   Step 6

   Ek deep dive chalao. Paid plan pe koi Research task chalao aur progress panel khol ke loop dekho. Free plan pe, kai sources chahne wala web-search sawal chalao aur citations check karo (Idea 8, khule mein)

7. 7

   Step 7

   Apne weekly workflow se ek skill banao, interview se ya "turn this chat into a skill" se. Draft review karo, enable karo, test karo ke wo fire hoti hai (Idea 5: expertise desk pe visit karti hai)

A.10 Kya Change Hota Hai, Kya Nahi

Poora appendix age karta hai. Model names badalte rehte hain, prices move karti hain, buttons migrate karte hain, features preview se default tak graduate hoti hain. Jab ye page aur live product disagree karein, product sahi hai — aur official Claude Help Center aur Anthropic ki prompt-engineering documentation current sources hain.

“Jo age nahi karta: har control jo aap kabhi milenge, kisi bhi product mein, 9 ideas mein se kisi ek ka handle hai — frozen weights ke beech ek selector, desk pe reasoning kitni jaye iska dial, sahi scope pe window mein text daalne ka mechanism, ya loop mein wired ek tool. Jab koi naya feature ship ho aur tutorials usay explain karne ki koshish kar rahe hon, ye test chalao jo ye course ne sikhaya: ye desk ka kaunsa tenant hai, ya loop ka kaunsa step hai? Usually tutorials se pehle aapke paas answer hoga.”

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.