Features build karna: foundation, phir AI chat
PL-4 aur PL-5 ko Jira se merged code tak le jayein. Architecture stage par inquisitive manager banna seekhein, tickets ke darmiyan context saaf rakhna, aur streaming ke baare mein aik real design debate jeetna.
Build karne ka waqt aa gaya. FeatureDev rails par laga kar aur sab kuch connect karke, hum apne pehle do product tickets — PL-4 (technical foundation) aur PL-5 (AI chat) — Jira se merged code tak le jate hain. Is page par aap yeh bhi seekhte hain ke architecture stage par inquisitive manager kaise banein, aur tickets ke darmiyan context ko saaf kaise rakhein.
01PL-4: foundation build karein
Aik line se shuru hota hai:
feature-dev:feature-dev implement Jira ticket PL-4 and make a PR.
Yeh PL-4 ko pull karta hai aur is baar process ko zyada precision ke sath follow karte hue, sharpe clarifying questions poochta hai — auth functional hona chahiye ya placeholder? (placeholder), NDA form client-side rakhna hai ya backend tak persist karna hai? (client-side). Phir yeh architecture-design stage tak pahunche kar rukta hai aur poochta hai: “kya yeh approach aage badhne ke liye theek lag rahi hai?”
Yeh review step aapka power ka moment hai — agar aap zyada technical bhi nahi hain to bhi. Aap “chat about this” choose kar sakte hain aur probe kar sakte hain: SQLite kyun? Single container kyun? Wohi questions poochein jo aik acha manager poochta hai taake woh trade-offs samajh sakein. Is process ke zariye guide hone par, Claude woh galtiyan nahi karta jo unguided projects mein ki thi (jaise aik monolithic main.py). Structure sensible lag rahi hai — approve karein aur build karne dein.
02Khud test karein, phir merge karein
Yeh report deta hai ke kaam ho gaya — 76 tests pass, API endpoints build, PR raised. Trust, but verify. Start script run karein aur open karein:
scripts/start-mac.sh # builds & starts the container # open http://localhost:8000 — type “New York”, watch the preview update, # download the PDF — all still working, now on real infrastructure
Yeh naye backend par chalta hai, live preview kaam karta hai, PDF download hota hai. Phir: “please merge the PR locally, push to main, and switch branch to main.” Foundation ship ho gaya.
03Tickets ke darmiyan context hygiene
Agle bade ticket se pehle, notice karein ke context almost full hai. Aik smart habit: Claude ko progress CLAUDE.md mein likhne dein, phir reset karein.
CLAUDE.md update karein1 min
“Add concise details to the end of CLAUDE.md on what has been implemented, and change anything that is no longer accurate.” Ab important state aik file mein hai, sirf conversation mein nahi.
Phir /clear karein10 sec
Conversation ko /clear se wipe karein aur agla ticket fresh shuru karein. Aapki CLAUDE.md (chhoti memory file, ~1.8k tokens) abhi bhi loaded hai — to kuch important loss nahi hua.
Log compact se sahi wajah darte hain: jab Claude full context ko squash karta hai, woh selectively forget kar sakta hai woh cheezein jo matter karti thin — kabhi kabhi CLAUDE.md ke hissay bhi — aur performance degrade ho sakti hai baad mein. Progress ko file mein likhna aur clean /clear karna compact par gamble karna se zyada safe hai.
04PL-5: AI chat add karein
Re-authenticate karein (yahan hang hona sirf auth reminder hai), phir: feature-dev:feature-dev implement Jira ticket PL-5 and make a PR. Iske clarifying questions excellent hain — bilkul woh jo aap product owner se poochte:
| Yeh poochta hai… | Hum choose karte hain |
|---|---|
| Chat UI form ko replace kare ya coexist? | Form ko completely Replace karein |
| Document preview: live ya submit par? | Live updates |
| Conversation kaise shuru ho? | AI greet kare aur pehla sawal pooche |
| Jab sare fields fill ho jayein? | Confirm karein aur download dikhayein |
Architecture stage par yeh teen specialized agents parallel mein launch karta hai — aik minimal, aik clean, aur aik pragmatic design — aur trade-offs samjhata hai. Yeh wahan aapka product knowledge kaam aata hai:
“Main simple backend chahta hoon, lekin sirf ek LLM call aur koi streaming nahi — Cerebras itna fast hai ke streaming ki zaroorat nahi. Ek structured-output call jo response aur extracted fields dono return kare woh cleaner hai.”
“chat about this” choose karne se review aik real conversation ban jata hai. Claude ne streaming ke liye reasonable case banaya; aap ne uski knowledge ke bahar ki baat samjhayi (Cerebras kitna fast hai), aur usne maan liya ke simpler design behtar hai. Woh back-and-forth — genuine engineering debate — exactly woh tareeqa hai jisse great results milte hain. Phir yeh jaakar implement karta hai.
✓ Aham baatein
- Ticket ko aik line se start karein:
feature-dev:feature-dev implement Jira ticket PL-4 and make a PR. - Architecture-review stage par, inquisitive manager banein — approve karne se pehle trade-offs probe karein.
- Hamesha khud result test karein merge karne se pehle, jab yeh claim kare ke sare tests pass hain.
- Tickets ke darmiyan: progress ko
CLAUDE.mdmein likhein, phir/clearkarein — auto-compact par bharosa karna se zyada safe, jo selectively forget kar sakta hai. - PL-5 ke liye, aik real design debate (ek structured call, koi streaming nahi kyunki Cerebras fast hai) ne cleanest architecture produce ki.