Find the perfect skill
CUDA Index Width in PyTorch
Development
Fix large-tensor indexing overflows in PyTorch CUDA kernels. Choose between 32-bit and 64-bit index math using canonical utilities like canUse32BitIndexMath and AT_DISPATCH_INDEX_TYPES.
Autonomous Feature Engineering Experiment
Data & AI
Run an autonomous feature engineering experiment loop to improve the statcast-gbm model by analyzing residuals, generating hypotheses, and testing candidates.
Durable Task Tracking with Beads
Development
Use Beads (bd) to manage shared project tasks, dependencies, and in-progress work. Learn the essential CLI workflow for durable cross-session tracking.
Global Validation
Security
Implements comprehensive server-side input validation with complementary client-side validation for user experience. Uses allowlists, type checking, and sanitization to prevent SQL injection, XSS, and other injection attacks. Apply when validating any external data entering the application, ensuring consistent security across all entry points and maintaining data integrity with field-specific error messages.
VideoDB: Video & Audio SDK
Development
Analyze, index, and edit video & audio. Capture, transcode, search, and generate subtitles, summaries, and real-time alerts.
Create a Pulumi event end to end
Content
Creates a Pulumi event end to end: scaffolds content bundle, generates social cards, files marketing issue, opens PR.
Next.js Performance Optimization
Development
Focuses on critical Next.js performance fixes: prevents asynchronous waterfalls via Promise.all and Suspense, reduces bundle size by banning barrel imports and using next/dynamic for heavy components, secures Server Actions with mandatory internal authentication, and enforces proper production build practices. Ideal when writing React components, data fetching logic, Server Actions, or optimizing bundle size in Next.js applications.
Confidence-based Model Tier Escalation
Development
Defines when to escalate agent tasks from Haiku to Sonnet to Opus based on confidence signals like test failures, clippy warnings, or repeated rejections. Enforces a never-skip-tiers rule and includes a diagnostic step to distinguish model capability issues from other problems. Use this skill when an agent fails or underperforms to systematically determine the correct tier.
End-to-End Development Workflow
Development
Executes a development task end-to-end: research, implement, CI, PR, and ship. Mostly autonomous with an optional plan-review pause.
Agent File Review
Development
Audit agent files: frontmatter, triggers, voice, length, duplication. Use when authoring or reviewing Claude agents.
Ask Multiple-Choice Question
Productivity
Present a multiple-choice question to the user using the AskUserQuestion tool. Collect a single answer from 2-4 options.
OpenRouter Context Optimization
Data & AI
Optimize context window usage for OpenRouter models to reduce cost and improve quality. Covers token estimation, conversation trimming, chunking, and prompt caching.