Context engineering & agents.md
Output mukammal taur par input par depend karta hai. Is input ko sahi karna — yani context ko — poori game hai. Yahan dekhein isme kya kya shamil hai, kam context kyun zyada behtar hota hai, aur agents.md file agent ko kaise steer karti hai.
01Context sab kuch kyun hai
LLM stateless hota hai: output sirf input par depend karta hai. Is input ko context kehte hain — woh tamam information jo model ke paas kaam karne ke liye hoti hai. Kyunki output input ka function hai, context ko sahi karna hi sab kuch hai.
Hum pehle isey prompt engineering kehte the. Ab yeh context engineering ban chuka hai — yeh samajhte hue ke sirf prompt nahi, balki model ke aaspas ki har cheez zaroori hai: tools, memory, files. Model ko sahi information dene ka poora toolkit.
02Context ki anatomy
Us single input mein bohot kuch jata hai. Roughly order yeh hai:
- System prompt — general framing: model ka role, job aur tone. Pehle aata hai.
- Tool descriptions — actions ka menu jo woh request kar sakta hai. Zyada tools = zyada space consumed.
- Memory — information jo conversations ke darmiyan persistent rehti hai.
- The conversation so far — sab se bara part. Har pichla message, reasoning, generated code aur tool calls. Sub kuch context mein squeeze kiya jata hai.
03Context window (aur kam space zyada kyun behtar hai)
Kitne tokens fit ho sakte hain iski ek hard limit hoti hai — context window. Exceed karne par call fail ho jati hai.
| Model | Context window |
|---|---|
| OpenAI GPT-5.2 | 400,000 tokens |
| Claude Sonnet 4.5 / Opus 4.5 | 200,000 tokens |
| Gemini 3 (Antigravity) | 1,000,000 tokens |
Sirf limit issue nahi hai. Limit se pehle bhi context bharne se quality degrade hoti hai — speed nahi, balki accuracy aur logic. Conversation ke start mein best results miltay hain jab context taqriban khali hota hai. Less is more.
Compacting — summariser
Limit cross hone par Claude Code jaise tools compacting run karte hain: conversation ko summarize kar ke purane messages replace kar dete hain. Yeh useful hai, par kabhi kabhi zaroori details drop ho sakti hain.
Purani aadat agent ko rok kar agents.md rewite karne ki thi. Lekin modern compacting bohot achi ho chuki hai. Trust karein, yeh relevant parts preserve rakhta hai.
04agents.md — agent ko guide karne wali file
agents.md file agent ko project ke baare mein instructions deti hai. Yeh ek markdown file (.md) hai jo simple aur human-readable hoti hai.
Just text, plus a little markup
# Heading H1 hai, ## Sub H2 hai, - bullet banata hai. Code ko backticks mein wrap karein.
Same idea, different filenames
Cursor / Codex / Copilot agents.md use karte hain. Claude Code claude.md use karta hai. Antigravity gemini.md. Concept wahi hai.
Hierarchy
Project root mein agents.md hamesha load hoti hai. Aap kisi subdirectory mein bhi rakh sakte hain — woh sirf tab load hoti hai jab agent us folder ki files par kaam karta hai. Inner file ki rules outer files ko override karti hain.
Root file hamesha context mein hoti hai. Subdirectory files tabhi include hoti hain jab agent us folder ko touch karta hai.
Achi agents.md mein kya hota hai
Isey prompt ki tarah likhein: concise, crisp, clear, maximum signal per word — kyunki yeh context space use karti hai.
- Overall project goals aur success criteria.
- Doosre documents ke links.
- Clear coding standards.
## Coding standards - Simpler is better. Never over-engineer; always simplify. - Comments only when necessary. Be concise. - Keep READMEs short. IMPORTANT: no emojis, ever. - Avoid over-defensive programming; handle exceptions only when needed. - Use latest, idiomatic library versions as of today.
Block capitals mein IMPORTANT likhna waqai help karta hai. Aur positive instructions dein — LLMs ko yeh yaad rakhne mein mushkil hoti hai ke kya nahi karna. Repeatation bhi thodi chal sakti hai.
Do tarah ki soch
agents.md par bohot mehnat
Success sharp agents.md se aati thi: root + subfolder files, plans aur criteria. Baar baar edit aur reset karna.
Let it hang out
Control agent par chodh dein. End goal par focus karein, skills, loops, sub-agents aur swarms use karein aur self-correct hone dein.
Serious projects ke liye 2025 mindset (context engineering) abhi bhi best hai, jab ke small projects ke liye 2026 approach fit hai.
✓ Aham baatein
- Output input par depend karta hai, is liye context engineering poori game hai.
- Context stack: system prompt, tools, memory, agents.md, aur poori conversation.
- Less is more: space bharne se quality degrade hoti hai.
- Modern compacting par bharosa karein.
agents.mdagent ko steer karta hai: concise, positive, IMPORTANT in caps.