A Claude Code plugin agent that intercepts vague prompts before execution. Detects your stack, asks targeted questions, and forges a structured prompt for Claude Code. No re-prompts. Zero wasted credits.
Works for every major stack: Next.js, React Native, Expo, Supabase, Prisma, FastAPI, Chrome MV3, Tailwind, shadcn, Clerk, Vercel AI SDK, LangChain, Drizzle, Cloudflare Workers, and 30+ more. Each stack has its own pitfall rules, scope anchors, and stop conditions baked in - so the forged prompt is never generic and common.
| Tool | What it does | Stars |
|---|---|---|
| prompt-master | A free Claude.ai skill that writes the accurate prompts for ANY AI tool. | 4,000+ ⭐ |
Requirements: Claude Code v2.0+ · Python 3.8+
Step 1 - Add the marketplace via terminal
claude plugin marketplace add nidhinjs/prompt-mini
Step 2 - Install via terminal
claude plugin install prompt-mini@nidhinjs-marketplace
Step 3 - Run this script
Mac / Linux:
python3 ~/.claude/plugins/cache/nidhinjs-marketplace/prompt-mini/0.1.0/setup.pyWindows:
python "$env:USERPROFILE\.claude\plugins\cache\nidhinjs-marketplace\prompt-mini\0.1.0\setup.py"Step 4 - Restart Claude Code
In Claude Code → Verify with /prompt-mini → you should see prompt-mini listed.
Running the script is required due to a known Claude Code bug where plugin hooks don't register automatically.
Step 1 - Clone
git clone https://github.com/nidhinjs/prompt-mini.gitStep 2 - Create a parent folder
Mac / Linux:
cd ..
mkdir prompt-mini-dev && cd prompt-mini-dev
mkdir .claude-plugin
echo '{"name":"local-dev","owner":{"name":"nidhinjs"},"plugins":[{"name":"prompt-mini","source":"./prompt-mini"}]}' > .claude-plugin/marketplace.json
mv ../prompt-mini ./prompt-miniWindows:
cd ..
mkdir prompt-mini-dev; cd prompt-mini-dev; mkdir .claude-plugin
[System.IO.File]::WriteAllText(".claude-plugin\marketplace.json", '{"name":"local-dev","owner":{"name":"nidhinjs"},"plugins":[{"name":"prompt-mini","source":"./prompt-mini"}]}')
mv ..\prompt-mini .\prompt-miniStep 3 - Register and install
claude plugin marketplace add /absolute/path/to/prompt-mini-dev
claude plugin install prompt-mini@local-dev
Step 4 - Run setup
python prompt-mini/setup.pyStep 5 - Restart Claude Code
Every prompt you type gets evaluated by a lightweight Python hook:
Clear prompt? → Passes through instantly. Zero questions.
Vague or framework-specific prompt? → Hook intercepts. Skill fires. Asks 2-6 targeted questions. Assembles a structured prompt and executes it.
You type a prompt
↓
Hook evaluates (~189 tokens)
↓
Clear? ──────────────────→ Pass through instantly
│
Vague?
↓
Scans project files (package.json, CLAUDE.md, imports)
Detects your stack automatically
Asks ≤5 clarifying questions
Assembles structured 6-block forged prompt
↓
Claude Code executes the forged prompt
Add * at the start of any prompt to skip prompt-mini entirely:
* fix the login bug
Other auto-bypasses: /slash-commands and #memory-commands pass through unchanged.
Framework specific routing, scope anchors, and stop conditions for every major stack.
| Stack | What prompt-mini prevents |
|---|---|
| Next.js App Router | Adds "use client" everywhere, mixes App/Pages Router, forgets "use server" on actions |
| Next.js Pages Router | Uses server actions (App Router only), wrong data fetching pattern |
| Remix | Confuses loaders with server actions, misses nested routing |
| SvelteKit | Writes React patterns in Svelte files, wrong file conventions |
| Astro | Adds React hooks to .astro files, treats it like a SPA |
| Vite + React | Attempts server actions, forgets all fetching is client-side |
| Vite + Vue 3 | Mixes Options API and Composition API |
| Nuxt 3 | Mixes Nuxt 2 patterns, wrong composables structure |
| Qwik City | Treats it like React, misses resumability patterns |
| T3 Stack | Creates REST routes instead of tRPC procedures, misses root router |
| Astro | Adds hooks to .astro files, forgets island component boundaries |
| Stack | What prompt-mini prevents |
|---|---|
| React Native + Expo | Uses DOM APIs, imports CSS, forgets platform file conventions |
| Flutter | Generates React patterns in Dart, forgets stateful vs stateless widget |
| Kotlin + Jetpack Compose | Mixes View system with Compose, wrong lifecycle handling |
| Swift + SwiftUI | Mixes UIKit patterns, wrong property wrappers |
| Capacitor | Treats it like a native app, forgets web bridge limitations |
| Stack | What prompt-mini prevents |
|---|---|
| Tauri 2 | Mixes Rust and frontend code, forgets to register commands in main.rs |
| Electron | Puts Node APIs in renderer, skips IPC bridge |
| Chrome MV3 | Uses background page (MV2), accesses DOM in service worker |
| Firefox WebExtension | Wrong manifest key names, missed permission declarations |
| VS Code Extension | Registers commands without adding to package.json contributes |
| Raycast Extension | Wrong API imports, misses required export structure |
| Wails | Mixes Go and frontend runtime assumptions |
| Stack | What prompt-mini prevents |
|---|---|
| FastAPI | Forgets Pydantic v2 syntax, skips response_model |
| NestJS | Creates services without injecting them via module providers |
| Hono | Uses Express req/res instead of Hono context |
| Node + Express | No error middleware, missing async error handling |
| Django | Mixes function-based and class-based view patterns |
| Flask | No blueprint structure, missing error handlers |
| Go + chi | Wrong middleware chaining, missing context passing |
| Rust + Axum | Wrong error types, missing trait implementations |
| Stack | What prompt-mini prevents |
|---|---|
| Vercel AI SDK | Mixes v3 and v4 APIs, exposes API keys client-side |
| LangChain.js | Uses deprecated v0.1 patterns in v0.2+ |
| LangGraph | Forgets to compile graph, skips state schema definition |
| Anthropic SDK | Puts API key client-side, wrong streaming pattern |
| OpenAI SDK | Wrong tool_choice format, misses response parsing |
| Mastra | Wrong agent configuration, misses tool registration |
| Stack | What prompt-mini prevents |
|---|---|
| Supabase | Uses client-side auth checks, disables RLS, wrong SSR client |
| Prisma | Schema changes without migration, forgets prisma generate |
| Drizzle | Mixes v0.28 and v0.29+ APIs |
| MongoDB | Wrong aggregation pipeline syntax |
| Redis / Upstash | Wrong command syntax for edge runtime |
| Turso | Uses Node-only libsql in edge environment |
| NextAuth v5 | Uses v4 session callbacks, wrong auth() usage |
| Clerk | Uses deprecated authMiddleware, forgets ClerkProvider |
| Supabase Auth | Client-side session checks, skips middleware refresh |
| Better-Auth | Wrong adapter configuration |
| Lucia | Outdated session handling pattern |
| Convex | Wrong mutation vs query pattern |
| Stack | What prompt-mini prevents |
|---|---|
| Tailwind CSS 4 | Generates tailwind.config.js (v3 pattern, removed in v4) |
| shadcn/ui | Generates component code instead of using CLI |
| Radix UI | Missing required accessibility props |
| MUI | Wrong theme provider nesting |
| Mantine | Missing MantineProvider wrapper |
| Vercel | Uses Node-only APIs in Edge Runtime |
| Cloudflare Workers | Uses Node APIs not available in Workers runtime |
| Fly.io | Wrong Dockerfile for the stack |
| Railway | Missing start command in package.json |
| Docker | Wrong base image for the runtime |
| Rule | What it does |
|---|---|
| Constraints in first 30% | Critical MUST NOT rules go at the top — they decay at the tail and get ignored under execution pressure |
| Real file paths only | Scope says src/lib/auth.ts not "in the auth module" |
| Specific stop conditions | Names exact destructive actions — never "be careful" |
| One task per prompt | Two distinct deliverables = Prompt 1 + Prompt 2, never merged |
| MUST and NEVER over weak words | "should not" → "MUST NOT". Weak words get ignored. |
| No ghost features | Nothing added beyond what was asked. Scope creep burns credits. |
| Checkpoint on every multi-step task | ✅ [what was completed] after each step — catches failures before they compound |
Does it slow me down? Clear prompts: no. The hook is a ~189 token Python check. Vague prompts: 5–15 seconds for the question flow. Faster than re-prompting twice.
Does it change my prompt without asking? No. It asks questions first. You answer. Then the forged prompt goes to Claude Code — you see it before anything executes.
What if I don't answer all the questions? It uses sensible defaults and notes the assumptions in the forged prompt's Context block.
Does it work for beginners? Yes. It detects your experience level from your prompt vocabulary. Beginners get simpler options. Senior devs get fewer questions.
Can I use it with Cursor / Windsurf? Claude Code only for now. Other tools on the roadmap.
- 0.1.0 — Initial release. Core hook + skill engine. 40+ stacks. 6 prompt templates. 35 credit killing patterns dectected.
MIT — see LICENSE for details.