Bade team codebases par kaam karna
Coding agents chote aur saaf projects par bohot zabardast kaam karte hain. Purani criticism yeh thi ke woh aik massive inherited codebase par phas jaate hain. Ab yeh baat kaafi had tak ghulat ho chuki hai — lekin abhi bhi kuch rules of the road hain jo aik bari team ko kamyab rakhte hain. Zyada tar baatein ab aapko common sense lagengi; yeh aik achha sign hai.
Yeh techniques wild new cloud tricks se zyada foundational (“2025”) skills par depend karti hain. Aik 50-bandon ki repo par, discipline humesha novelty ko beat karta hai.
01Bade codebases kyun alag hote hain
Aik chote from-scratch project par, agent poori cheez apne zehen (context) mein rakh sakta hai, docs likh sakta hai, aur zero-shot demo ko nail kar sakta hai. Aik bohot bare codebase ko inherit karna mushkil hai: wahaan context mein aane se kahin zyada code hota hai, bohot se log use change kar rahe hote hain, aur aik laparwah instruction kafi nuqsan pohncha sakti hai. Last year se models isme kafi behtar ho chuke hain — lekin neeche diye gaye practices hi ise sahi chalaye rakhte hain.
021 — Apne agent docs mein invest karein
Aapki AGENTS.md / CLAUDE.md files sab se bara lever hain. Trick yeh hai ke progressive disclosure follow karein: directory tree ke har level par documentation ho, jo sirf tab reveal ho jab agent us subdirectory mein enter kare.
/backend/market_data mein enter hota hai woh us folder ki doc padhta hai aur fauran interface jaan leta hai — poori repo load karne ki zaroorat nahi.Zyada detail context waste karti hai; bohot kam detail ho to ghulat functions call ho jaate hain. Tight docs ka aim rakhein jo package ke interface ko reflect karein. Aap Claude se unhe write karwa sakte hain — lekin phir unhe check karein, review rounds karein, aur code change hone par docs ko revise karne ka step apne process mein rakhein.
032 — Docs ko link karein, embed nahi
Jab aap @ tag se kisi document ko reference karte hain, to uski tamam contents outer document mein insert ho jaati hain — jo aam taur par aap nahi chahte, aur yeh aapka context khatam karne ka tez tareen rasta hai.
Poori file dump kar deta hai
See @docs/market-data-spec.md tamam 1,400 lines ko context mein paste kar deta hai chahe unki zaroorat ho ya na ho.
Agent ko choose karne dein
“Market-data spec (interface + providers): docs/market-data-spec.md.” Agent isko sirf tabhi read karta hai jab task ko zaroorat ho.
Is liye docs ko aise structure karein ke summarise karein, phir point karein detailed documents ki taraf. Is tarah aik agent bari repo ko efficiently navigate karta hai bajaye usme doobne ke.
043 — Aik consistent team workflow
Aik agreed process pick karein aur poori team ko us par align karein. Agar har koi alag tarike se karega to khichdi ban jayegi — Jira tickets kaun file karega? @claude kaun tag karega? Aik single flow (jaise Jira → FeatureDev → PR, ya Claude ke liye tagged GitHub issues) ka matlab hai ke har kisi ko pata hai ke naye feature ko order karne ka sahi tarika kya hai.
054 — Plugins aur skills ko favour karein
Pehle plugins
Apne project ke liye sahi plugins layein — e.g. code-simplification, FeatureDev — aur unhe shared process ka hissa banayein taake har koi aik hi tarike se kaam kare.
Domain skills
Is baat ke liye skills likhein ke aapka project kaam kaise karta hai — aapke frameworks, aapke patterns (e.g. market-data API use karne ka sahi tarika). Yeh tamam agent work mein common standards ko majboot karte hain.
065 — Robust, na ke brittle, tests
Aik majboot test suite humesha zaroori raha hai (80% coverage kafi arse tak gold standard tha). Lekin coverage percentage par fixate na hon — LLMs iske peeche bhaagte hain aur heavily mocked, brittle tests likhte hain jo real behaviour ke bajaye har code path ko check karte hain.
Aise tests jo re-implementation ke baad bhi survive karein lekin jab logic waqai break ho to fail ho jayein. Apni testing strategy par feedback dein (ise AGENTS.md mein dalein), aur jab bhi aap “testing ke khatir” over-mocked tests ka dher dekhein to push back karein — bilkul waise hi jaise aap kisi human ke liye karte.
076 — Quality ka owner human hai
Aakhir mein, code ke liye aap hi accountable hain. Agent sirf aik tool hai. Aik human reviewer ke sath aisa culture banayein jahan baat aakar rukti ho, aur agent ke “slop” ko reject karne ki pakki aadat dalen: zaroorat se zyada lambi files, over-defensive code, ya review karne mein takleef dene wali koi bhi cheez.
Ab bohot saara code generate karna bohot aasan ho gaya hai — aur sara bojh human par chala jata hai ke woh is sab ko review kare, jo ke sakht mehnat ka kaam hai. Is liye insist karein ke agents succinct (muhtasar aur samajh aane wala) code likhein jo reviewer ko overwhelm na kare, aur loop mein disciplined review ko qaim rakhein.
087 — Chote bite-sized chunks
Aik massive project par, Claude ko kabhi “refactor the whole codebase” bol kar tag na karein — aap khud apne liye musibat ko dawat de rahe hain. (Aap hamare capstone jaise chote project par to yeh kar sakte hain, lekin 50-bandon ki repo par nahi.) Iske bajaye, kisi human ko — ya khud Claude Code ko — bare kaam ko chote steps mein divide karne dein, jisme se har step independently specifiable, testable aur human-reviewable ho, aur phir unhe aik aik karke execute karein.
09Aik chota assignment
Apni industry se koi popular open-source project clone karein. Agent se kahein ke code mein koi TODO dhoond kar usko poora kare, aur dekhein woh kaisa kaam karta hai. Tree ke har level par detailed AGENTS.md banayein, FeatureDev plugin try karein, aur in practices ko use karein. Anti-test ke taur par, 10 iterations ke liye Ralph loop mein “refactor/simplify everything” try karein — aur dekhein kitna badnuma result aata hai. Yeh contrast aapko sikhata hai ke kya kaam karta hai aur kya nahi.
✓ Aham baatein
- Progressive-disclosure docs at every subdirectory are the biggest lever — keep them tight and updated.
- Link to docs, don’t
@-embed them — let the agent decide what to read. - Agree on one team workflow; standardise with plugins and domain skills.
- Write robust tests, not brittle over-mocked ones; don’t worship coverage %.
- A human owns quality — reject slop, demand succinct code, and work in bite-sized chunks.