◆ Vibe EngineeringSMIT · AI Agentic Engineering English
Class 10 · 10.1

Building day: SaaS blueprint

Week two ka mukammal hona — aik blue building day. Pre-Legal SaaS product banane se pehle, hum bunyad rakhte hain: skills waqai mein kaise kaam karti hain, aur do qism ke CLAUDE.md, jismein aik lean global bhi shamil hai jo har session mein load hota hai.

🎯 Goal: build ke liye bunyad⏱ Read: 14 min🧭 Type: concept + setup

Week two ke mukammal hone mein khush aamdeed — week two, day five. Yeh aik blue day, aik building day hai, aur iske khatam hone tak humare paas aik working SaaS platform hoga: humara legal-document drafter, Pre-Legal. Kal humne chalana seekha Claude Code ko aik disciplined workflow ke zariye. Aaj hum us skill ko aik asli product par point karte hain aur isko ticket by ticket banate hain. Pehle, hum bunyad sahi tareeqe se rakhte hain — skills, aur do qism ke CLAUDE.md.

01Hum kahan hain — aur wo aik cheez jo humne skip ki

Aik aakhri, quick recap big three ka. MCP dosron ke tools ko jor deta hai (bara ecosystem, lekin context-hungry aur kabhi kabhi flaky — Jira auth yaad hai). Skills simple Markdown hain jo expertise efficiently add karte hain. Plugins yeh sab kuch bundle kar dete hain. Kal ki build ne aik Jira ticket MCP ke zariye parhi, isko FeatureDev plugin se guzara, aur aik GitHub PR ship kiya — sab kuch aik command se.

Gap ko pehchano

Humne MCP aur aik plugin use kiya… lekin kabhi actually aik skill likhi nahi. Aaj hum yeh theek karte hain — hum apni khud ki custom skill banate hain aur product ko isko use karne dete hain. To pehle, aayein waqai samajhte hain ke skill kya hai.

02Skill waqai mein kaise kaam karti hai

Aik skill ka khubsurat design hai: progressive disclosure. Sirf aik chota sa hissa hamesha loaded rehta hai; baqi on demand parha jata hai, to yeh aapke context ko tabhi touch karta hai jab zaroorat ho.

1 · Metadata (always read) name + one-line description 2 · Instructions (read if relevant) the body of SKILL.md 3 · Extra files & scripts (only if needed) The model decides how deep to read Always sees level 1. Reads level 2 only if the task matches. Opens level 3 files/scripts only when it actually needs them. context-friendly by design
Teen levels of depth. Metadata sasta hai aur hamesha mojood; baqi sab lazily load hota hai.

Metadata aur instructions aik saath aik file mein rehte hain, SKILL.md, metadata upar special format mein hota hai (hyphens se fence kiya hua block — aur detail agle page par). Skills simple file-system architecture use karti hain: aik .claude folder (aapke project aur/ya aapki home directory mein) jismein andar skills/ folder hai. Wahan correctly-structured skill folder daal dein aur Claude Code isko use kar sakta hai. Yeh pura mechanism hai.

03Do CLAUDE.md files hain

Kisi bhi real project ke shuru mein aap aik acha CLAUDE.md likhte hain. Lekin yahan aik cheez hai jo humne pehle nahi kahi: aik se zyada hain — aur sirf wo nahi jo aap subfolders mein scatter kar sakte hain.

Project CLAUDE.md

📁 Repo mein rehti hai

Is project se juri: yeh kya hai, isko kaise banana hai, iski conventions. Team ke sath committed aur shared.

Global CLAUDE.md

🏠 Aapki home directory mein rehti hai

Aapse juri, kisi project se nahi. Har Claude Code session mein loaded, har jagah.

Jaise aap apni home .claude folder mein skills rakh sakte hain, waise hi aap wahan CLAUDE.md bhi rakh sakte hain. Yeh aapki personal standing instructions hain — aapka style, aapki pet peeves — aapke Claude ke sath tamam kaam par lagu.

04Aik lean global CLAUDE.md likhna

Kyunke yeh har context mein load hota hai, golden rule yeh hai: isko chhota rakhein, aur har lafz count kare — kyunke har token count karta hai. Anthropic pehle se hi Opus mein bohot sari achi defaults bake kar chuka hai, to isko sirf un extra nuggets ke liye use karein jo aap hamesha chahte hain. Aik mufeed starting shape:

~/.claude/CLAUDE.md
# IMPORTANT — be simple
Approach tasks in a simple, incremental way. Small, simple steps.
Validate and check each increment before moving on.
Do not over-engineer. Do not program defensively.

# Code style
Use UV as the Python package manager: always `uv run`, never `python3`;
always `uv add`, never `pip install`.
Favor clear, concise docstring comments; be sparing elsewhere.
Favor short modules. Keep the README concise. Clean up old files.
Never use emojis in code, print statements, or logs.

# Debugging
Always identify the root cause before fixing — prove it, do not guess.
Reproduce consistently. Change one thing at a time. Be methodical.
Do not jump to conclusions. Do not apply workarounds.
Isko apna banayein

Bas isko copy na karein. Shayad aapko emojis ya lambi READMEs pasand hain — inko rakhein. Shayad aap vanilla Python aur pip ko UV par prefer karte hain. Point yeh hai ke apni non-negotiables ko aik dafa encode karein, takay har session usi tarah shuru ho jaise aap chahte hain. Dekhein ke debugging lines bilkul kal ki playbook hain, kuch teez reminders mein distill ki gayi.

Aik cheez jo beginners ko uljhati hai

Aapki home directory mein .claude folder Claude Code ki apni machinery rakha hai (telemetry, project info) — usko haath na lagayein. Aap sirf apna khud ka CLAUDE.md add kar rahe hain (aur, optionally, skills/) iske saath.

✓ Ahm points

  • Aaj hum aik asli SaaS product banate hain, Pre-Legal — aur aakhir kar apni khud ki skill likhte hain.
  • Skills progressive disclosure use karti hain: metadata hamesha parha jata hai; instructions aur extra files sirf zaroorat par load hoti hain. Yeh sab .claude/skills/ ke neeche SKILL.md mein rehta hai.
  • Do CLAUDE.md files hain: aik project wali repo mein, aur aik global wali aapki home directory mein jo har session mein use hoti hai.
  • Global wali ko chhota rakhein — har token count karta hai; sirf apni hamesha-sachi preferences capture karein.
  • Ache staples: simple/incremental rahein, over-engineer na karein, package-manager rule, emojis nahi, aur pehle root-cause debugging.