Official AI dev assets

12 extremely opinionated Cursor rules that enforce zero-tolerance TypeScript. No `any` (ever), strict null checks, exhaustive switch statements, branded types for IDs, const assertions, discriminated unions for state machines, and proper error types. Every rule includes a BAD example and a GOOD example. For developers who are tired of `as any` and `// @ts-ignore`.

15 Cursor rules (.mdc files) optimized for minimal token consumption while maintaining full behavioral control. Every rule includes a measured token cost annotation, and most are under 100 tokens. The set includes before/after comparisons showing how typical verbose rules waste tokens on examples the model already knows. Covers: code style, error handling, testing, TypeScript, React, API design, git workflow, security, and accessibility. Designed for developers on Cursor Pro who hit usage limits, or anyone who wants more context window for their actual code.

A four-agent review team for Claude Code: Security Reviewer (OWASP + CWE, auth/authz, injection, secrets), Performance Reviewer (N+1 queries, bundle size, memory leaks, rendering), Accessibility Reviewer (WCAG 2.2 AA, keyboard nav, screen readers, contrast), and Test Coverage Analyzer (untested paths, edge cases, mutation testing gaps). Each agent has a focused checklist, specific prompt template, and structured output format. Designed to replace ad-hoc PR review checklists with consistent automated review.

A lightweight persistent memory system for Claude Code that uses CLAUDE.md files and structured markdown memory files to maintain project knowledge, decisions, and session context across sessions. Unlike claude-mem (which requires SQLite + ChromaDB + an MCP server), this approach uses zero dependencies — it's pure file-based, git-friendly, and works on any machine Claude Code runs on. Includes memory schema, auto-save hooks, retrieval patterns, cleanup strategies, and honest token cost analysis.

A Claude Code skill with agent definition templates, orchestration patterns, error recovery strategies, and a decision framework for when to use managed agents vs subagents. Includes YAML agent definitions, Python session management code, coordination patterns for multi-agent workflows, and honest token/cost breakdowns. Built for developers who have API access and want to ship autonomous agents without re-learning infrastructure.

A production SKILL.md for Claude Code that orchestrates four specialized agents in a structured code review and testing pipeline. The architect agent validates design decisions, the reviewer catches bugs, the test-writer generates comprehensive tests, and the security scanner finds vulnerabilities. Runs as a single /review-pipeline command with parallel execution and consolidated reporting.

Pre-configured MCP server setup for PostgreSQL (query + migrate), Redis (cache management), GitHub (PRs + issues + actions), Sentry (error tracking + triage), and Vercel (deploy + env management). Complete JSON configs for Claude Code and Cursor with connection string templates, auth setup, permission scoping, and usage examples for each server.

18 opinionated Cursor rules for production Next.js 16 apps. Covers `use cache` directive with cacheLife/cacheTag, Partial Prerendering (PPR), middleware stacking for auth + i18n + rate limiting, Server Actions with optimistic updates, ISR revalidation strategies, error boundary patterns, Turbopack config, and a pre-deploy checklist. Not generic advice — specific patterns with code.

Ten production-grade Claude Code hook presets for `~/.claude/settings.json` or a plugin's `hooks/` directory. Covers all nine public hook events: PreToolUse, PostToolUse, UserPromptSubmit, Stop, SubagentStop, SessionStart, SessionEnd, PreCompact, Notification. Each hook wraps with `timeout`, reads `tool_input` from stdin JSON (not just `CLAUDE_FILE_PATH`, which breaks in 2.1.x+), uses the correct exit code semantics (0 allow, 2 block with stderr feedback), and is idempotent. Based on the official hook reference, the anthropics/claude-code plugin-dev skill, and bugs filed in issues #13744, #24327, #19009, #40580.

Cursor rules for implementing Stripe Checkout in a Next.js App Router project. Every rule addresses a real, documented gotcha: raw-body handling in route handlers, signature verification, idempotency against the 3-day retry schedule, the `checkout.session.completed` vs `invoice.paid` race, and the middleware body-parser trap. Based on the canonical stripe-node/nextjs example plus six issues filed against vercel/next.js and vercel/nextjs-subscription-payments.

A Claude Code skill for Prisma 7 + PostgreSQL workflows. Enforces the migrate dev vs migrate deploy split, the DATABASE_URL vs DIRECT_URL split for pooled providers, PgBouncer transaction-mode compatibility, the Vercel postinstall caching fix, N+1 detection, and the shadow-database failure mode. Triggers on P2002/P2025/P1001/P1008/P1017, on schema edits, and on the exact error strings users paste from Vercel build logs and Prisma CLI output.

Two working project-instruction templates (AGENTS.md + CLAUDE.md) for a Next.js App Router monorepo and a Python FastAPI service. Scoped under the 5k-token guidance from HumanLayer's 'Writing a good CLAUDE.md', compatible with the agents.md 1.0 spec (nearest-file-wins), and designed around the bugs actually filed against OpenAI Codex (issues #6502, #8601, #3198) where the agent reads AGENTS.md and then ignores it anyway.

Claude Code skill (progressive disclosure) that triggers on TS error codes and error-message fragments. The base SKILL.md stays under 800 tokens; `references/error-catalog.md` carries the full 30-error matrix with DO/DON'T code blocks; `references/type-narrowing.md` and `references/generics.md` load on demand. Based on actual filed issues in microsoft/TypeScript, angular/angular, vercel/next.js, and the typescript-eslint rule set — not on vibes.

Opinionated Cursor rules for React 19 (released December 2024), grounded in actual React repo issues and react.dev docs. Covers the `use()` hook (and its Suspense fallback quirk from facebook/react#30701), `useActionState` (the renamed `useFormState` — facebook/react deprecated the old name), `useOptimistic` (including the rollback bugs from facebook/react#31944, #31967, and the Next.js revert-before-revalidate issue vercel/next.js#49619), form `action` props, `useFormStatus`, `ref` as a plain prop (forwardRef is deprecated — codemod `react/19/remove-forward-ref` exists), the React Compiler's effect on manual `useMemo`/`useCallback` (facebook/react#31913, #31924), the `use client` / `use server` network boundary and its serialization rules, and Document Metadata hoisting with its deduplication gap. No "use TypeScript" fluff. Every rule has a source.

Cursor rules for using shadcn/ui the way its own maintainers describe it: copy-paste source code you own, Radix primitives underneath, CSS variables for theming, Tailwind utilities via cn(). Covers the Tailwind v4 / OKLCH / @theme-inline migration, the cn() + twMerge contract, the wrap-nothing composition rule, the Form + react-hook-form + Zod triad, and the update-sync problem that happens when you customize too aggressively. Based on canonical patterns from ui.shadcn.com plus real bugs in github.com/shadcn-ui/ui issues #6427, #6585, #6856, #10101, and Hacker News discussions on the copy-paste model.