Class 10 · 10.5

AI legal-doc generator ko test karna

Demo ko asli engineering se alag karein: PR parhein takay confirm ho ke aapki skill use hui, AI chat ko user ki tarah test karein (aur Cerebras ki speed mehsoos karein), feedback ko fixes mein tabdeel karein, aur PL-6 mein har document type ship karein.

🎯 Goal: verify aur expand⏱ Read: 14 min🧭 Type: test + build

PL-5 ban gaya hai — aik AI chat jo user ko interview karta hai aur document fill karta hai. Ab hum wo do cheezain karte hain jo demo ko engineer se alag karti hain: code parhein takay confirm ho ke usne humari skill use ki, aur isko user ki tarah test karein. Phir hum seedha PL-6 par chalte hain aur har document type ko support karte hain.

01PL-5 ne kya deliver kiya

Claude success report karta hai: aik AI chat interface, backend API, LiteLLM jo Cerebras se baat kar raha hai, structured outputs, aur live preview updates — sab kuch takriban 13 files aur 456 lines ke PR mein. Paper par impressive hai. Lekin pehle do checks aate hain.

02Kya actually humari skill use hui?

PR kholein aur code parhein. Hum Cerebras skill ke fingerprints dhoondh rahe hain jo humne likhi thi:

Dhondhein

Aik Pydantic schema

Fields ke liye structured-output model — bilkul wahi jo humne manga tha.

Pehchan

`extra_body`

Mojood hai — bilkul usi tarah liya gaya jaise humne skill mein likha tha. Saboot ke skill parhi aur use ki gayi.

Dhondhein

Sahi model

gpt-oss-120b OpenRouter ke zariye, Cerebras par pinned.

Yeh woh lamha hai jab skills kaam aati hain

Apni khud ki extra_body aur provider routing ko generated code mein verbatim dekhna pura point hai: aapne expertise aik dafa encode ki, aur Claude ne isko wafadari se apply kiya — aur har teammate ke liye bhi yahi karega jo repo clone karega.

03Isko user ki tarah test karein

Server start karein aur localhost:8000 kholein. Aik greeting nazar aati hai: “Main aapki madad karunga aik mutual non-disclosure agreement banane mein. Chalo basics se shuru karte hain…” Iski sawaalat ka jawab dein aur dekho ke dayen haath ka preview fill ho raha hai.

Speed mehsoos karein

Responses fauri tor par wapis aate hain — pehla chunk basically pura jawab hai. Yeh Cerebras hai, aur yeh bilkul wahi wajah hai ke humne kaha streaming zaruri nahi. Aik zabardast demonstration deliberate design choice ka jo kaam kar gaya.

Yeh perfect nahi hai, aur yeh theek hai — testing hai building. Do imaandaar feedback pieces samne aate hain: chat kabhi kabhi sawal poochna band kar deta hai aur conversation sukh jati hai (isko hamesha agla sawal poochna chahiye agar zyada zaroorat ho), aur aap sirf aakhir mein pahunch kar download kar sakte hain. Hum inko note karte hain — yeh agli ticket mein fixes ban jate hain. Isko 90% victory kahein: asli AI chat, OpenRouter par, humari nayi skill use karti hui.

04Merge karein, record karein, reset karein

Wahi rhythm jaise pehle: PR ko locally merge karein aur main par push karein; yaqeeni banayein ke CLAUDE.md nayi status reflect karti hai; /clear; aur agli ticket ke liye Atlassian re-authenticate karein. Chhoti, repeat hone wali discipline — yeh wahi hai jo multi-ticket build ko mush mein tabdeel hone se bachati hai.

05PL-6: har document type — aur wo fixes

PL-6 product ko tamam templates tak expand karta hai jo humne curate kiye. Aur kyunke hum kar sakte hain, hum feedback ke do tukre usi request mein fold kar dete hain:

claude

implement Jira ticket PL-6. Also please make a couple of fixes:
after answering a question, return focus to the text input field,
and always ask a follow-on question if more information is needed.
Test, then make a PR.

Yeh PL-6 complete kar deta hai, do fixes apply karta hai, apni khud ki review se kuch quality improvements add karta hai, aur aik PR raise karta hai. Isko test karein: Cloud Service Agreement ka pucho — yeh type ko fauri tor par pehchan leta hai, sawaalat ki saaf run poochta hai (company names, dates, governing state, subscription period, pricing, payment terms), aur dayen taraf preview ko live fill karta hai, sahi brand colours mein. Download PDF press karein — aik mukammal, polished document. Aik aur 2,500-line PR, aik aur feature ship ho gaya.

Feedback loop hi workflow hai

Dekho ke yeh kitni qudrati tareeqe se flow karta hai: build → user ki tarah test karo → concrete feedback capture karo → isko agli ticket mein fold karo. Aap code nahi likh rahe; aap steer kar rahe hain — woh notice kar rahe ho jo aik asli user karega, aur isko precise instructions mein tabdeel kar rahe ho.

✓ Ahm points

PR par bharosa karne se pehle, code parhein — yahan hum extra_body, Pydantic schema, aur sahi model confirm karte hain, saboot ke humari skill use hui.
User ki tarah test karein. Cerebras responses ko instant mehsoos karata hai — “no streaming” decision ko validate karta hai.
Asli feedback (chat ka sukh jana; download timing) failure nahi — yeh agli ticket mein precise fixes ban jata hai.
Rhythm rakhein: merge → CLAUDE.md update → /clear → re-auth.
PL-6 tamam document types aur do fixes aik instruction mein add karta hai — steering, coding nahi.