Ye Notes Kya Hain
Ye page Panaversity ke Agent Factory Foundations course ke saare chapters ka ek detailed aur asaan Roman Urdu revision guide hai. Maqsad simple hai: mid-term quiz se pehle har chapter ka core concept, real-world application, aur exam-ready recap ek hi jagah mil jaye, bina complicated English jargon ke.
Har chapter mein teen cheezein hamesha milengi. Pehla, ek Core Idea box jo us poore chapter ka essence ek jagah deta hai. Doosra, detailed explanation practical examples ke saath, taake concept sirf ratta na lage balke samajh mein aaye. Teesra, ek recap table jo revision ke waqt ek nazar mein sab yaad dila de. Aakhir mein ek cheat sheet aur 10 sawalon ka self-test quiz bhi hai.
Ye guide un logon ke liye bhi useful hai jo AI agents, chatbots, ya automation systems par kaam kar rahe hain, kyunke har concept ke saath ek practical engineering angle bhi diya gaya hai, na sirf exam ka nazariya.
Chapter 00
Orientation, The AI-Native Company Model
10-80-10 Rule, Digital FTE, aur is poore course ka roadmap
Core Idea
Kaam AI era mein teen layers mein hota hai. Pehle aap ek general agent use karte hain problem solve karne ke liye, phir specialized AI Workers banate hain repeatable jobs ke liye, phir un Workers ko mila kar ek AI-Native Company banate hain jahan human sirf direction aur verification deta hai.
Har Kaam Insaan Se Shuru Hota Hai
Chahe kitna bhi advanced automation system ho, har professional engagement ek human se shuru hoti hai jo ek general agent ko direct karta hai. Sawal sirf ye hota hai ke kaunsa agent chuna jaye, aur ye poori tarah is baat par depend karta hai ke aap actually achieve kya karna chahte hain. Ye poori philosophy ek simple lekin powerful formula par khari hai jisay 10-80-10 Rule kehte hain.
1 · General Agent
Problem solve karne ke liye
2 · Specialized AI Workers
Repeatable jobs ke liye
3 · AI-Native Company
Human sirf direction aur verification deta hai
10-80-10 Rule Ka Matlab
Pehle 10%, Human Intent
Clear prompt, spec ya goal set karna. Sabse zyada leverage isi stage mein hai, ye ghalat hui to baaki 90% bhi galat direction mein jayega.
Beech ke 80%, AI Execution
AI heavy lifting karta hai: summarizing, drafting, generating, analyzing, formatting. Time yahan sabse zyada bachta hai.
Aakhri 10%, Human Verification
Quality check, output sharp karna, final approval. Ye stage kabhi skip nahi hoti, chahe AI kitna bhi confident lage.
Practical Example
Digital FTE, Sirf Ek Prompt Nahi
Digital FTE (Digital Full Time Employee) ka matlab sirf ek achha model ya achha prompt nahi hai. Ye ek poora system hai jo char cheezein combine karta hai:
Domain Expertise
Aapka apna specialized knowledge, jis field mein aap kaam karte hain
Explicit Specifications
Documented rules aur instructions, jaise spec file ya brief
Engineering Architecture
Proper tools, memory aur workflow design
Human Oversight
Verification loop jo kabhi khatam nahi hoti
Course Ka Roadmap
Prerequisites sequence yehi hai jo ye poora course follow karwata hai, taake koi bhi shortcut na le aur foundation strong rahe:
| Concept | Ek Line Mein |
|---|---|
| 3 Layers | Agent, phir Worker, phir AI-Native Company |
| 10-80-10 Rule | Intent set karo, AI se karwao, phir khud verify karo |
| Digital FTE | Expertise + Spec + Architecture + Oversight, chaaron zaroori |
| Course Roadmap | Thesis se Foundations se Mode-specific courses tak |
Chapter 01
What AI Actually Is, A Crash Course
Prediction machine, tokens, context window, hallucination, aur agents ki asal mechanic
Core Idea
AI ek "next token predictor" hai, librarian nahi. Ye fact search nahi karta, sirf itna predict karta hai ke agla piece of text kya aana chahiye. Iske paas sach check karne ka koi internal organ nahi hai. Sab kuch isi ek fact se nikalta hai.
1. Predicts, Lookup Nahi Karta
Jab aap poochte hain "France ki capital kya hai", AI kisi database mein France to Paris search nahi karta. Wo sirf itna predict karta hai ke "The capital of France is..." ke baad sabse plausible continuation kya hai, aur training data mein "Paris" itni baar aaya hota hai ke wahi predict hota hai. Common facts pe prediction aur lookup same result dete hain, isliye farq nazar nahi aata. Lekin jab topic rare ho, tab AI ke paas koi "sach" continue karne ko nahi hota, to wo sabse plausible-sounding cheez bana deta hai. Wo lying nahi kar raha, uska yehi kaam hai: continue karna, chahe sach ho ya na ho.
Practical Example
2. Training Ek Dafa Hui, Phir Freeze Ho Gayi
Do terms yaad rakhein. Training ek dafa hoti hai, company ke paas, model banate waqt; ismein weights (numbers) set hote hain. Inferencehar dafa jab aap use karte hain; wahi frozen weights chalte hain, kuch change nahi hota. Jab aap chat mein AI ko correct karte hain aur wo "haan aap sahi hain" kehta hai, wo seekh nahi raha, sirf ek plausible reply predict kar raha hai. Naya chat kholein, wahi purani ghalti dubara aayegi. Isi wajah se knowledge cutoff hota hai, aur isi wajah se AI ko aapka private data pata nahi hota, kyunke wo kabhi training text mein tha hi nahi.
3. Koi Second Faculty Nahi Jo Sach Check Kare
Insaan ke paas do faculties hoti hain: ek jo jawab generate karti hai, dusri jo check karti hai "kya mujhe yakeen hai iska". AI ke paas sirf pehli faculty hai. Wahi mechanism jo sahi jawab banata hai, wahi ghalat bhi banata hai; koi internal flag nahi hota farq batane ke liye. Yehi hallucination hai. Ye bug nahi hai, ye machine ka exactly wahi kaam hai jo wo design se karti hai: plausible continuation, chahe sach ho ya na ho.
4. Ye Letters Nahi, Tokens Padhta Hai
Text pehle tokensmein chop hota hai: chunks, usually ek word ya word ka hissa. "Strawberry" jaise word ko wo 2-3 chunks mein dekhta hai, letters individually nahi. Isi wajah se AI kabhi kabhi "strawberry mein kitne R hain" jaisa simple sawal bhi ghalat count kar deta hai.
Practical Example
5. Context Window Hi Uski Poori Duniya Hai
Weights frozen hain, koi apni memory nahi, to model sirf wahi dekh sakta hai jo context windowmein maujood ho: aapka prompt, conversation history, uploaded files, system prompt. Isko "reading desk" samjhein, brain nahi. Jo cheez desk pe nahi rakhi, wo model ke liye exist hi nahi karti, chahe aapko kitna bhi obvious lage. Isi wajah se lambi conversations mein quality girti hai: purani cheezein desk se hat jati hain ya summarize ho jati hain.
Engineering Angle
6. Confidence Ek Learned Style Hai, Sach Ka Proof Nahi
Training ke baad models ko human feedback se tune kiya jata hai jisay RLHF kehte hain. Log confident, agreeable jawabon ko zyada rate karte hain, chahe wo sahi ho ya na ho. Isliye model confident sound karna seekh leta hai as a style, aur sycophancybhi isi se aati hai, yani aapse agree karne ki tendency. Fix yehi hai ke neutral framing use karein, jaise "iske dono sides evaluate karo", ya score maangein: "1 se 10 scale par grade karo".
7. Jagged Frontier, Ek Jagah Brilliant, Agli Jagah Useless
Insaan ki ability smooth hoti hai: agar koi hard calculus kar sakta hai to easy arithmetic bhi kar lega. AI ki nahi. Wo legal contract clause likh dega perfectly, aur "strawberry" mein letters miscount kar dega. Farq training data ki frequency se aata hai: jo tasks common thay unme wo strong hai, jo rare hain unme weak hai.
Practical Rule
8. Tools Use Karke Ye Act Karta Hai
Pure text predictor sirf text de sakta hai, real duniya mein kuch kar nahi sakta. Tools (web search, code execution, file read, API call) is limit ko todte hain. Mechanism simple hai:
Yehi definition hai "agent"ki: same predictor, plus tools, plus loop, jo goal ki taraf repeat hota hai. Koi naya "mind" nahi hai, sirf predictor + tools + loop. Claude Code ke subagents bhi isi loop ka scaled version hain.
9. "Thinking" Bhi Bas Extra Prediction Hai
Reasoning models pehle apna kaam (steps, working) predict karte hain, phir usi ko apne context mein rakh ke final answer predict karte hain. Ye sach mein help karta hai kyunke reasoning desk pe rakhne ke baad answer predict karna asaan ho jata hai. Lekin isse dusri faculty nahi milti; reasoning bhi wahi single process hai jo galti kar sakta hai. Zyada thinking gap kam karta hai, khatam nahi karta.
| # | Idea | Ek Line Takeaway |
|---|---|---|
| 1 | Predicts not lookup | Fluency sach nahi, plausibility hai |
| 2 | Training frozen | Knowledge cutoff, private data blind |
| 3 | No truth-checker | Hallucination normal operation hai |
| 4 | Tokens not letters | Non-English costlier in tokens |
| 5 | Context window = duniya | Jo desk pe nahi wo exist nahi |
| 6 | Confidence = style | Sycophancy isi se aati hai |
| 7 | Jagged frontier | Easy task bhi fail ho sakta hai |
| 8 | Tools = action | Agent = predictor + tools + loop |
| 9 | Thinking = extra prediction | Gap kam karta hai, khatam nahi |
Chapter 02
AI Prompting in 2026
Retrieval modes, context rot, sycophancy fix, brainstorm-iterate loop
Core Idea
Har advanced prompting technique asal mein sirf do moves hain: sahi context andar daalna, ya galat context bahar rakhna. Baaki sab isi ka variation hai.
Novice Vs Power User
Novice "which car should I buy" pooch ke generic jawab leta hai. Power user spec sheets, insurance quotes, apni driving pattern ka data upload karta hai, phir "trade-offs batao, think hard" bolta hai. Mental model yaad rakhein: AI ek highly motivated fresh grad hai jo aapke baare mein kuch nahi janta. Jitna aap usay brief karenge utna behtar output aayega.
Teen Retrieval Modes
Pretrained
Fast lekin stale. Static sawal ke liye theek hai.
Web Search
“Latest on X” jaisi cheezon pe trigger hota hai. Model original page nahi padhta; retrieval layer ka condensed summary milta hai, isi se summary drift aata hai. Sources specify karein, exact quote maangein.
Deep Research
Heaviest mode. Minutes leta hai, dozens of sources scan karta hai, structured report banata hai.
Practical Example
Talking to AI Ka Real Mechanic
System prompt aapko nazar nahi aata lekin har chat mein already load hota hai. Aap apni personal instructions bhi add kar sakte hain; ye exactly wahi cheez hai jo CLAUDE.md files mein hoti hai subagents ke liye.
Context rot ek real problem hai. Ek lambi conversation mein multiple unrelated topics mix karna performance girata hai. Chat lambi hone par tools chupke se purani baaton ko compact kar dete hain: summary bana ke original detail replace kar dete hain. Rule: jab topic change ho, naya chat kholein. Agar kuch save karne layak hai, pehle file mein save karein, phir reset karein.
Reasoning mode("think hard") ab explicitly invoke ki ja sakti hai. Simple lookups pe mat use karein, slow aur costly hai. Complex multi-input decisions pe zaroor use karein.
Sycophancy Ka Fix Mechanical Hai
In verbs se bachein
find, defend, confirm, prove: AI conclusion pehle se maan ke chalta hai
Ye verbs use karein
evaluate, compare, critique, find any: neutral framing, honest jawab
Sabse powerful move: number maangein."Is ye code sahi hai" ke bajaye "har criterion ko 1-10 grade karo, justification ke saath." Adjectives ("strong", "solid") aapko decide karne layak kuch nahi dete, numbers dete hain.
Brainstorm-Iterate Loop
Ye is chapter ka sabse high-leverage habit hai. Seedha final draft mat maango:
Practical Example
Text Se Aage, Aur Safe Use
- Image input coarse detail dekhta hai, fine detail par weak hai.
- Data analysis mein hamesha confirm karein ke AI actually code run kar raha hai, guess nahi. “Write and run code, show me the code” explicitly bolein.
- Desktop apps (Cowork, OpenWork) plan-review-approve workflow follow karte hain. Delete kabhi bhi recycle bin mein nahi jata; permission hamesha smallest scope se start karein.
- Model selection jagged hai, koi ek best nahi. Har mahine leaderboard check karein aur apna common task 2-3 models mein try karein.
- Models checking models sabse high-stakes technique hai: ek model se self-critique karwayein, high-stakes decisions par doosri model family se bhi grade karwayein. Dono ke beech disagreement hi asli signal hai jahan blind spot chhupa hai.
| # | Concept | Practical Takeaway |
|---|---|---|
| 1 | Novice vs power user | Brief AI jaise naye colleague ko |
| 2 | 3 retrieval modes | Wording se mode trigger hoti hai |
| 3 | Context window/system prompt | Naya topic, naya chat |
| 4 | Sycophancy | Verbs badlein, number maangein |
| 5 | Brainstorm-iterate loop | Pehle options, phir hi expand |
| 6 | Data analysis | Code run hote dekhna zaroori hai |
| 7 | Models checking models | Cross-family disagreement asal signal hai |
Chapter 03
Markdown In, HTML Out
Structure ki asymmetry, spec skeleton, document format decisions
Core Idea
Agent ko likhte waqt Markdown use karein, agent se jawab mangte waqt HTML mangwayein. Decision hamesha ek sawal se hoti hai: ye output last mein kaun padhega.
Teen Jagah, Teen Format
Teesri row sabse important hai. Jab aap ek chat ka context doosre chat mein copy karte hain, wo bhi "agent to agent" hai, chahe dono side aap hi baithe hon. Wahan Markdown rahegi, HTML nahi. Test hamesha yehi hai: agar insaan browser mein padhega, HTML mangwayein. Agar AI ne dubara padhna hai, Markdown mein rakhein.
Markdown Ka Poora Syllabus, Sirf Paanch Cheezein
- Headings importance dikhate hain. Ek document mein ek hi title, level skip mat karein, aur heading ko label mat rakhein, claim banayein. “Budget” ki jagah “Budget: PKR 50,000 hard ceiling” likhein.
- Bullets vs Numbers: bullets ka matlab set hai, order matter nahi karta. Numbers ka matlab sequence hai, order hi instruction ka hissa hai.
- Triple backtick fences batate hain “ye data hai, instruction nahi”. Error message, example output, ya kisi aur ka quote fence ke andar rakhein.
- Links: jab aap URL prompt mein dete hain, AI asli page visit kar ke padh sakta hai; summary se guess karne ki jagah asli source use hota hai.
- Images: bracket ke andar wala description hi wo cheez hai jo AI dekhta hai. Isay caption samjhein jo batata hai kis cheez pe focus karna hai.
Spec Skeleton
Ye woh structure hai jo real client projects mein use hoti hai:
01
Goal
02
Context
03
Requirements
04
Hard Constraints
05
Out of Scope
06
Expected Output
Do sections sabse zyada kaam karte hain. Out of Scope agent ke sabse common failure ko rokti hai: over-delivery. Expected Output format drift ko rokta hai.
High-Leverage Habit
HTML Kyun Mangwayein
Test simple hai: kya aap ye poora plain text padhenge? Agar nahi to HTML mangwayein. HTML mangwate waqt 4 cheezein zaroor batayein: kaun padhega, kya include ho, interactive chahiye ya nahi, aur kaise padha jayega.
Paanch HTML patterns jo sabse zyada kaam aate hain:
- Decision grids: options cards mein, trade-off label ke saath
- Explainer reports: long document ko ek page summary mein
- Code review: color-coded diffs, annotated code
- Design prototypes: live sliders jab words se describe karna mushkil ho
- Throwaway editors: ek baar ke decision ke liye drag-drop tool
Social Media Aur Document Formats
WhatsApp/LinkedIn/Facebook plain text hain, formatting strip ho jati hai. HTML sirf link preview card aur designed images (PNG export) ke liye kaam ki hai. Document format sawal "insaan is output ka karega kya" se decide hota hai:
Sign / Print
Edit
Word
Present
Slides
Numbers
Excel
Tool feed
CSV
Key Rule
| Concept | Ek Line Takeaway |
|---|---|
| Direction asymmetry | Kaun last mein padhega, wahi decision hai |
| Headings/lists | Heading = claim, bullets = set, numbers = sequence |
| Spec skeleton | Build se pehle grade aur fix karein |
| HTML brief | Kaun, kya, interactive, kaise padhega |
| Social feeds | Plain text body, HTML sirf preview/image ke liye |
| Documents | Sign=PDF, Edit=Word, Present=Slides, Numbers=Excel |
Chapter 04
Code You Never Write
VPRF test, five-section brief, verification ladder, blast radius safety
Core Idea
AI ab sirf answer nahi deta, code likh ke run bhi kar deta hai, apne sandbox mein. Aap client hain, AI developer hai.
VPRF Test
Ye test decide karta hai koi task "code problem" hai ya sirf "answer problem":
Volume
Hath se karne layak se zyada items hain
Precision
Galti ki cost hai
Repetition
Ye kaam dobara hoga
Files
Data files mein rehta hai
Practical Example
Commissioning Discipline
Precision-critical kaam mein hamesha explicit bolein: "write and run code, show me the code you ran, pehle exact row count aur column names batao." Ye teesri line lie-detector hai: agar row count galat aaya to samajh jayein AI ne file actually padhi hi nahi.
Five-section brief replace karta hai casual prompting ko: Goal, Input, Output, Rules, Edge cases. Rules wahi jagah hai jahan aapka domain knowledge jata hai. Edge cases wo jagah hai jahan aap explicitly bolte hain blank/duplicate/corrupt data ka kya karna hai, warna AI khud guess karega aur wo guess silent rahega.
Verification Ladder
- 1
Known-answer test
Chota slice jiska jawab pehle se pata ho, us pe test karein
- 2
Reality check
Row count in vs out, basic sanity numbers
- 3
Plain-English replay
AI se poochein step by step logic batao; galat logic English mein bhi galat lagega
- 4
Adversarial pass
“Apni analysis mein galti dhoondo” bolein
- 5
Cross-model check
High-stakes cases mein doosre model se bhi verify karwayein
Errors Aur Reusability
Errors dialogue hain, failure nahi. Red error poori paste kar dein, AI khud diagnose kar leta hai. Agar number galat lag raha ho lekin error na aaye, symptom report karein expected value ke saath.
Keep the script:ek dafa solve hua problem ko script + brief.md pair bana ke folder mein rakhein. Agli baar sirf "isi script ko naye data pe chalao" bolna kaafi hai.
Five Surfaces
Chat sandbox zero-risk, temporary, one-off jobs ke liye. Terminal agents (Claude Code, OpenCode) folder ko directly dekh sakte hain, script permanent rehti hai, error khud fix ho jati hai. Desktop apps (Cowork, OpenWork) plan-then-approve built-in rakhte hain. Rule of thumb: jab upload karna annoying lagne lage, wahi signal hai terminal/desktop surface pe move karne ka.
Blast Radius Rules, Production Safety
- Copies pe kaam karein jab tak script trusted na ho jaye
- Destructive action (rename/delete/move) se pehle dry run maangein; poori list dekhein approve karne se pehle
- Scope smallest folder tak rakhein, kabhi poori drive point mat karein
- Output naye file mein likhwayein, original ko kabhi overwrite mat karwayein
Client Work Angle
Edge of the Map
| Concept | Ek Line Takeaway |
|---|---|
| VPRF | Volume, Precision, Repetition, Files: ek bhi fire ho to code problem |
| Five-section brief | Rules aur Edge cases sabse zyada kaam karte hain |
| Verification ladder | Known-answer test kabhi skip mat karein |
| Keep the script | Brief + script + sample ek folder mein |
| Blast radius | Copy, dry run, scope, new output file |
Chapter 05
Skills and Connectors
Recipe vs kitchen analogy, SKILL.md anatomy, security checklist
Core Idea
Chat message ek dafa ka order hai, Skill har baar wahi kaam sahi tareeke se karne ka tareeka hai, Connector AI ko haath deta hai aapke real apps tak pahunchne ke liye.
Kitchen Analogy
Connector = Kitchen
Stove, chaqu, stocked pantry: yani Google Drive, Gmail, Slack, aapka tracker. AI ko haath deta hai aapke real apps tak pahunchne ke liye. Kitchen bina recipe ke improvised aur inconsistent output degi.
Skill = Recipe Card
Recipe card jo batati hai dish aapke restaurant ke tareeke se kaise banti hai: har baar wahi kaam sahi tareeke se. Recipe bina kitchen ke sirf padhi ja sakti hai, cook nahi ki ja sakti.
Skill Technically Kya Hai
Ek folder jismein ek SKILL.md file hoti hai. Us file ke top pe do cheezein hamesha loaded rehti hain: name aur description. Neeche jo bhi likha hai wo tab tak load nahi hota jab tak description match na ho. Isay progressive disclosure kehte hain; isi wajah se aap dus-bees skills install kar sakte hain bina AI ko slow kiye.
Sabse Zaroori Baat
Connector Technically Kya Hai
Ek MCP server jo aapke app se safe connection banata hai. Teen facts yaad rakhein:
- AI aapki hi permissions inherit karta hai: jahan aap khud nahi ja sakte wahan AI bhi nahi ja sakta.
- Aap khud decide karte hain read-only ya read-write. Hamesha read-only se start karein.
- Har conversation mein alag se enable karna parta hai: connect karna aur enable karna do alag steps hain.
Farq Yaad Rakhne Ka Tareeka
| Feature | Kab Active | Kaam |
|---|---|---|
| Project | Hamesha on | Standing context/persona |
| Skill | On-demand fire | Specific task ka tareeka |
| Custom Instruction | Har jagah apply | Global preference |
| Connector | Per-chat enable | Real app tak access |
Sath Mein, Real Power Yahan Hai
Pattern simple hai: Connector real data fetch karta hai, Skill usay aapke tareeke se shape karta hai. Misal ke taur par, agar koi content generation system Google Drive se past posts pull kare (Connector) aur phir ek fixed brand format mein dhale (Skill), wo poora automation ek sentence mein ho sakta hai.
Kaunsa Chahiye, Teen-Step Test
- Friction ye hai ke “main baar baar explain kar raha hun kaise karna hai”: Skill chahiye.
- Friction ye hai ke “main baar baar doosre app se data copy-paste kar raha hun”: Connector chahiye.
- Dono ho to dono chahiye.
Skill Banana, Koi Code Nahi Likhna
skill-creator naam ki built-in skill hai jo aapke liye SKILL.md khud likh deti hai. Build loop:
Portability: Skill ek open standard hai, isliye ek jagah likha hua SKILL.md Claude.ai, Cowork, Claude Code, OpenCode, Codex CLI, aur Gemini CLI tak bhi chal jata hai. Lekin ChatGPT ke Custom GPTs aur Gemini ke Gems vendor-locked hain: ek jagah se doosri jagah portable nahi.
Security Checklist
- Trusted sources se hi skill install karein
- Enable karne se pehle SKILL.md khud padhein ya AI se padhwayein
- Connectors read-only se start karein, sirf zaroori scope tak access dein
- Poori drive kabhi mat connect karein
| Concept | Ek Line Takeaway |
|---|---|
| Kitchen analogy | Connector = kitchen, Skill = recipe |
| SKILL.md anatomy | Name + description hamesha loaded |
| Description | Yehi decide karti hai fire hogi ya nahi |
| 3-step test | Re-explain=Skill, copy-paste=Connector |
| Portability | Skills open standard, GPTs/Gems vendor-locked |
| Safety | Read before enable, read-only start, small scope |
Final
Quick Revision Cheat Sheet
Agar sirf 5 minute milein exam se pehle, sirf ye 6 lines dobara parh lein, poora course wapis yaad aa jayega.
Orientation: Agent se Worker se AI-Native Company; 10-80-10 rule sab kuch drive karta hai.
What AI Actually Is: AI predictor hai, truth-checker nahi. Context window hi uski duniya hai.
Prompting 2026: Sahi context andar daalo ya galat context bahar rakho, yahi har technique ka core hai.
Markdown In, HTML Out: Agent likhne ko Markdown, insaan ke padhne ko HTML.
Code You Never Write: VPRF test se decide karo code problem hai ya nahi, phir five-section brief se commission karo.
Skills and Connectors: Skill = kaise karna hai, Connector = kahan se data lena hai.
Self-Test
Khud Se Poochein: 10 Sawal
Pehle khud jawab dein, phir sawal pe click kar ke answer check karein. Agar 8+ sahi hain to aap ready hain.
1AI “France ki capital Paris hai” kaise jaanta hai, agar wo lookup nahi karta?
2Ek chat mein AI ko correct karne ke baad, doosri chat mein wo galti dobara kyun karega?
3Hallucination ko “bug” kehna kyun galat hai?
4Sycophancy fix karne ke liye kaunse do practical tareeke hain?
5Brainstorm-iterate loop ke 4 steps kya hain?
6Markdown aur HTML ka use kis sawal se decide hota hai?
7VPRF test ke chaar letters kya represent karte hain?
8Verification ladder ke 5 steps kya hain?
9Skill aur Connector mein bunyadi farq kya hai?
10Ek skill ki description sabse zyada important kyun hoti hai?
Downloads
PDF Notes Apne Paas Rakhein
Offline revision ke liye dono PDFs download karein: detailed notes poori tayari ke liye, quick revision exam se theek pehle ke liye.