AI aur Agent Tools

Agent ke saath kuch banane ka matlab hai dher saari chhoti configuration files likhna, jo chupchaap fail hoti hain. Aisi skill jiska description kabhi kisi task se match nahi karta; aisa permission rule jo wo command cover hi nahi karta jo aap chahte the; aisa hook jiska matcher ek aise tool ka naam leta hai jo hai hi nahi; aisa tool schema jo validate to ho jaata hai par phir bhi galat tareeke se call hota hai. Ye tools us kaam ko pehle hi jaanch lete hain, isse pehle ki wo aapka ek poora debugging session kha jaaye. Skill files, subagent definitions aur slash commands likhiye aur lint kijiye. MCP server entries, hooks aur plugin manifests banaiye, aur permission rules ko asli commands par test karke dekhiye ki kaunsa rule jeetta hai. Prompt ki taraf: rows ki table par template bhariye, do versions ko shabd-dar-shabd token antar ke saath milaiye, system prompt mein contradictions dhoondhiye, aur aise few-shot examples banaiye jo apne hi tags se bahar na nikal sakein. Aas-paas ke data ke liye: JSONL padhiye aur validate kijiye, fine-tuning dataset jaanchiye, streaming response ko wapas jodiye, aur document ko aise chunks mein baantiye jo asli seemaon par tootein. Aur kharch ke liye: context window ka budget banaiye, request ko aise kram mein rakhiye ki cache hone wala prefix jitna ho sake utna lamba ho, aur agent loop ka kharch turn-dar-turn nikaaliye, us history ke saath jo wo har baar dobara bhejta hai. Sab kuch aapke device par chalta hai — prompts, datasets aur configuration files kabhi upload nahi hote.

• Skill Builder — SKILL.md likhiye, frontmatter saath-saath check hota hai.
• Skill Linter — SKILL.md paste kijiye aur dekhiye use kisi kaam se judne se kya rok raha hai.
• Subagent Builder — Subagent ki definition file banaiye, uske tools aur model samet.
• Slash Command Builder — Arguments aur tools list ke saath apna command banaiye.
• Project Instructions Builder — Aisi CLAUDE.md likhiye jo har session mein apni jagah kamaye.
• Plugin Manifest Builder — plugin.json banaiye aur uski apekshit folder banawat dekhiye.
• Hooks Builder — settings.json ke hooks banaiye aur kabhi na chalne wale matchers pakadiye.
• Permission Rule Tester — Allow, ask aur deny rules asli commands par aazmaiye.
• MCP Server Config — stdio, HTTP ya SSE ke liye MCP server entry banaiye.
• MCP Tool Schema Builder — Tool tay kijiye aur JSON Schema paiye, strict mode jaanch ke saath.
• Tool Schema Checker — Definition paste kijiye aur jaaniye wo kyun reject ho rahi hai.
• Prompt Template Filler — Ek prompt ko rows ki table par failaiye aur kami dekhiye.
• System Prompt Linter — Contradictions, bekaar instructions aur bina bhare placeholders dhoondhiye.
• Prompt Diff — Do versions shabd-dar-shabd milaiye, token antar ke saath.
• Few-shot Example Builder — Input-output jodon ko tag, JSON ya chat roop ke examples banaiye.
• XML Prompt Section Builder — Sections ko aise tags mein lapetiye jinse content bahar na nikle.
• JSONL Viewer aur Validator — JSONL file line-dar-line padhiye, chhaantiye aur jaanchiye.
• Chat Transcript Converter — Messages array ko padhne layak transcript mein badliye, aur wapas.
• Dataset Validator — chat roop ka JSONL dataset un errors ke liye jaanchiye jo training gira dein.
• SSE Stream Viewer — Kacchi stream se text, tool calls aur usage wapas paiye.
• Context Window Planner — Window ko prompt, tools, history aur jawab mein baantiye.
• Prompt Cache Planner — Blocks aise kram mein rakhiye ki cache hone wala prefix sabse lamba ho.
• Text Chunker — Document ko asli seemayon par tootne wale overlap tukdon mein baantiye.
• Agent Loop Cost Estimator — Turn-dar-turn kharch nikaaliye, dobara bheje history aur cache ke saath.
• Batch Request Builder — Rows ki table ko alag pehchaan wale bulk job mein badliye.

Aksar pooche jaane wale sawaal

Ye AI tools kis kaam ke hain?

Model ko call karne ke liye nahi, balki uske aas-paas ki cheezein banane aur debug karne ke liye: skill aur subagent files, MCP server entries, hooks, permission rules, prompts, tool schemas, datasets aur cost estimates. Ye wo checks hain jo kisi configuration file ke live session mein chupchaap fail hone se pehle chala lene chahiye.

Kya inmein se koi AI tool mere prompts kahin bhejta hai?

Nahi. Is category ka har tool text processing hai jo aapke browser mein chalta hai. Prompts, datasets, transcripts aur configuration files aapke device par hi rehte hain — aur ye zaroori hai, kyunki in files mein aksar internal instructions aur customer data hota hai.

Kya model prices aur context limits up to date hain?

Har aankde ke saath wo tareekh dikhti hai jab use aakhri baar check kiya gaya tha, aur jo tools in rates se hisaab lagate hain unmein rates edit kiye ja sakte hain. Published prices badalte rehte hain, isliye built-in table ko shuruaati point maaniye aur budget banane se pehle provider ke apne pricing page se mila lijiye.

Kya ye AI tools sirf ek hi provider ke saath kaam karte hain?

File formats — skills, subagents, slash commands, hooks, MCP — un formats ke apne conventions follow karte hain. Baaki sab provider-neutral hai: JSONL, chat transcripts, streaming events, chunking, token budgets aur prompt structure sab ek jaise kaam karte hain, chahe aap unhe kisi bhi model ko bhejein.

Skill ya tool description itni dhyaan se kyun check karein?

Kyunki wahi ek field tay karti hai ki wo cheez kabhi use hogi ya nahi. Dhundhle description wali skill bina padhe padi rehti hai; patle description wala tool galat arguments ke saath call hota hai. Dono bina kisi error message ke fail hote hain — aur yahi wo problem hai jise ek linter achhe se pakadta hai.