Skills: capabilities add karne ka ek aasan tareeqa
MCP ke heavy weight ka ek lightweight solution. Markdown instructions with optional scripts, jo progressive disclosure ke zariye intelligently load hote hain — yeh clean kyun lagte hain, folder structure kaise kaam karta hai, aur inke drawbacks kya hain.
MCP powerful toh hai par heavy hai. Skills baad mein ek lightweight, clean alternative ke tor par samne aaye. Yeh kya hain, yeh itne clean kyun hain, aur inki kya limitations hain, aayein samajhte hain.
01Skill kya hoti hai
Ek skill Markdown mein likhi gayi instructions hoti hain — bohot hadd tak CLAUDE.md jaisi, lekin teen important differences ke sath: aap ke paas bohot sari skills ho sakti hain, inhe banana bohot aasan hai, aur yeh context mein intelligently load hoti hain. Yeh Claude Code ka concept tha jo ab doosre tools mein bhi phail raha hai.
Bohot saari CLAUDE.md files jaisi
Har specific kaam ke liye focused instructions ka ek folder.
Sirf files
Markdown aur optional scripts. Koi external process ya plumbing nahi — koi bhi bana sakta hai.
Progressive disclosure
Pehle sirf ek chota header load hota hai; baaqi content tabhi aata hai jab model ise relevant samjhta hai.
02Progressive disclosure — efficiency ki trick
Is mushkil naam ke peeche ek simple idea hai: har cheez ek sath load mat karein. Ek skill ke teen levels hote hain, aur Claude utna hi aage jata hai jitni zaroorat hoti hai.
Skill mein shell/Python scripts shamil ho sakti hain jinhe Claude run kar sakta hai. Yeh JSON signature wala formal tool call nahi hai — yeh sirf script run karke output read kar leta hai. Iska faida yeh hai ke processing context se bahar rehti hai; sirf final result wapas aata hai.
03Architecture: sirf folders hain
Is system ki beauty iski simplicity mein hai. Ek skill sach mein ek folder hota hai.
.claude/
└─ skills/
└─ my-great-skill/
├─ SKILL.md # metadata + instructions (required)
├─ reference.md # level-3 resource, loaded on demand
└─ scripts/
└─ do_thing.py # runnable, result-onlyDo placement options: project ka .claude (sirf is repo ke liye) ya aap ki home directory ka .claude (har jagah ke liye). Skill install karna sirf is folder ko copy karna hai — jiska matlab hai ke ise repo mein commit karke poori team ke sath share kiya ja sakta hai.
--- name: juggling description: Teaches Claude how to juggle. Use when the user asks to juggle. --- # Instructions Freeform guidance, workflows, code snippets… aur level-3 resources ke references.
04Skills kyun aachi hain — aur inki 2 kamzori¥an
Simple & context-light
MCP jaisa process overhead nahi. Build aur share karna aasan, by design efficient. Simple use-cases ke liye perfect hai.
Kam power, fuzzy discovery
Rich function signatures nahi hain — sirf simple inputs wale scripts hain. Aur discovery keyword-matching par depend karti hai.
Skill trigger hogi ya nahi, yeh aap ke prompt ki wording par depend karta hai jo top metadata se match honi chahiye. Aap "use" keh kar force kar sakte hain, lekin agar exact name na pata ho toh yeh mushkil ho sakta hai.
Bohot se developers skills ko Claude Code equip karne ka behtareen tareeqa samajhte hain. Aaj ka rule yeh hai: skills ko MCP par prefer karein, siwaye wahan jahan MCP majboori ho (e.g. Massive for market data) ya jahan koi tool dono support karta ho.
✓ Aham baatein
- Skill = Markdown instructions (+ optional scripts); aap bohot saari create kar sakte hain.
- Progressive disclosure: metadata pehle, instructions match hone par, resources zaroorat par.
- Scripts context se bahar run hote hain aur sirf result return karte hain.
- Yeh sab
.claude/skills/ya~/.claudemein folders hote hain; Git ke zariye easily share hote hain. - Pros: simple, context-light. Cons: MCP jaisi full power nahi. Jahan ho sake skills prefer karein.