Class 11 · 11.2

Foundations recap: machine asal mein kaise kaam karti hai

Pehle hafte ka sab se qeemti din “heavy theory” ka din tha — kyunke aik baar jab aap samajh jaate hain ke model dar-haqeeqat kya kar raha hai, to aage aane wali har trick jadu lagna band ho jaati hai. Yeh usi bunyad ka khulasa hai.

🎯 Goal: theory ko refresh karna⏱ Read: 15 min🧠 Type: revision

Hafton ki tooling sab neeche mojood model ke chand bunyadi concepts par tiki hui hai. Agar in mein se koi bhi cheez thori kamzor lage, to yeh wohi page hai jahan thehar kar samajhna hai — finale mein har cheez inhi par mabni hai.

01LLM asal mein kya karta hai

Tamam hype ko hata dein to aik large language model dar-asl aik next-token predictor hai. Yeh ab tak ke text ko padhta hai, usay tokens (characters ke tukron) mein todta hai, aur agle sab se zyada imkani token ki prediction karta hai — baar baar. Bas yahi hai. Har chalaak kaam jo yeh karta dikhai deta hai, woh isi kaam ko intehai shandar tareeqe se anjam dene se nikalta hai.

Koi memory nahi, koi plan nahi — sirf is waqt samne mojood har cheez ki bunyad par agla most-likely token.

Memory ka illusion

Model aapke pichle message ko yaad nahi rakhta. Har turn par, poori conversation dobara input ke tor par bheji jaati hai. Jo cheez memory lagti hai woh asal mein har martaba poora transcript dobara padhna hai. Yeh akela fact context windows, compacting, aur is baat ki wazahat karta hai ke agents.md itni zyada ahmiyat kyun rakhti hai.

02Agent banane wale chaar tricks

Aik raw model sirf text predict karta hai. Chaar cheezein is autocomplete ko aik aise agent mein badal deti hain jo dunya mein kaam kar sakta hai:

1 · Reasoning

Jawab dene se pehle sochein

Model ko pehle apni soch likhne dein — yeh multi-step results ko zabardast tareeqe se behtar banata hai.

2 · Tools

Bahar ki dunya se rabta

Isay aise functions dein jinhein yeh invoke kar sake — file padhna, command chalana, web search karna. Yahi woh beej hai jahan se MCP ugta hai.

3 · The loop

Act, observe, repeat

Tool chalayein, result wapis feed karein, agla qadam tay karein — jab tak task mukammal na ho loop karte rahein.

4 · Autonomy

Goal ko own karna

Isay script nahi balki objective dein, aur steps chune ka ikhtiyar dein. Yahi tool se agent tak ki chhalang hai.

Aik precise tareef

Aik agent dar-asl loop mein chalne wala LLM hai, tools ke sath, jo aapke tay kardah goal ki taraf kaam karta hai. Claude Code bilkul yahi hai: reasoning + aik toolbox (files read/write karna, shell chalana, search) + aik loop + aik goal. Sub-agents aur swarms — finale mein — isi tarah ke mazeed loops hain jinhein hoshiyari se arrange kiya gaya hai.

03Context engineering

Kyunke model har turn par sab kuch dobara padhta hai, is liye context window (iski working memory) aapka sab se qeemti resource hai. Is mein kya jata hai usay manage karna context engineering kehlata hai, aur yahi woh skill hai jo pareshan-kun sessions ko smooth sessions se alag karti hai.

Lever	Yeh kya hai	Kab use karein
system prompt	Woh mustaqil hidayat jo agent hamesha dekhta hai.	Aik baar bunyadi qawaneen set karne ke liye.
/context	Dekhein ke is waqt window kitni bhari hui hai.	Jab jawabat ghair-wazeh ya slow hone lagein.
/compact	Space bachane ke liye history ka khulasa (summary) banayein.	Task ke darmiyan, jab window bhar jaye.
/clear	Naye task ke liye bilkul saaf shuruat karein.	Kisi ghair-mutaliqa kaam par switch karte waqt.

Yeh finale mein kyun ghoonjta hai

Sub-agents bunyadi tor par main context window ko protect karne ke liye wajood mein aaye hain: woh apna shor-sharaba wala kaam apne alag context mein karte hain aur sirf jawab wapis laate hain. Agar context engineering aapko samajh aa gayi hai, to Class 12.4 mein sub-agents bilkul natural lagenge.

04agents.md & CLAUDE.md

Is poore course ki sab se zyada asar-angez (highest-leverage) file. Aik agents.md (ya Claude Code ke liye CLAUDE.md) plain-markdown brief hai jo automatically context mein load ho jaati hai — aapke project ke standing orders.

CLAUDE.md

# Project: Pre-Legal SaaS
## Stack
Next.js (frontend) · FastAPI (backend) · SQLite (db)

## Conventions
- Use uv run, never bare python
- All docs live in planning/
- Keep changes small; one feature per PR

## How to run
docker compose up  # opens on localhost:3000

Isay lean rakhein

Aik aisi CLAUDE.md jo sab kuch kehne ki koshish karti hai, aakhir mein kuch bhi nahi keh paati — yeh sirf context zaya karti hai. Mukhtasar, specific aur sachi batein lambi aur khwahishat par mabni baton se behtar hain. Humne Class 10.1 mein lean global CLAUDE.md ke sath yeh dekha tha.

05Boss banna — paanch usool

Class 5 se, woh aadatein jo vibe coding ko waqai kamyab banati hain — har paimane par sach sabit hoti hain:

Pehle spec banayein

Pehle wazeh taur par batayein ke aap kya chahte hain is se pehle ke woh aik line bhi likhe. Ghair-wazeh input, ghair-wazeh output.

Simple shuru karein

Sab se chhota version chalayein, phir usay expand karein. Mega-prompt ke lalach se bachein.

Chhote steps mein iterate karein

Build karein, dekhein, theek karein. Chhote loops galatiyon ko tab pakad lete hain jab unhein theek karna aasan hota hai.

Challenge karein

Poochhein “kya yeh mazeed simple ho sakta hai?” Agar aap chhor dein to woh khushi se over-engineer kar dega.

Frustration handle karein

Jab phans jayein to context reset karein, rewind karein, ya goal dobara bayan karein — daayire mein behes na karein.

✓ Aham nukat

Aik LLM next-token predictor hai; “memory” har turn par poora transcript dobara padhna hi hai.
Aik agent = model + reasoning + tools + loop + goal. Week 3 mein har cheez inhi loops ka izafa hai.
Context mehdood hai: /context, /compact, /clear use karein — aur sub-agents isay protect karne ke liye banaye gaye hain.
Aik lean CLAUDE.md aapki sab se highest-leverage file hai. Chhoti aur sachi file jeet jaati hai.
Boss banein: spec likhein, simple shuru karein, iterate karein, challenge karein, aur frustration ko pur-sukoon tareeqe se handle karein.