Skill UI

This guide provides a structured approach to red-teaming Large Language Models (LLMs) to detect sensitive data leakage from system prompts. It simulates advanced attacks—including prompt injection, instruction override, and encoding tricks—to identify embedded secrets, API keys, and proprietary business logic. Essential for validating adherence to OWASP LLM07 and securing AI applications against data exfiltration.

A comprehensive red-teaming skill for assessing the security posture of Retrieval-Augmented Generation (RAG) pipelines. It systematically probes two critical injection surfaces: poisoned retrieved context (indirect prompt injection) and embedding manipulation. Use this skill to validate retrieval guardrails, ensure data leak prevention, and demonstrate vulnerabilities in LLM-powered knowledge assistants.

This guide provides a comprehensive reference on Agent Skills, a standardized and powerful method for providing structured context, workflows, and tools to Large Language Models (LLMs). It details the anatomy of the SKILL.md file, the three-stage progressive disclosure loading process, and best practices for creating both capability (tool-based) and process (workflow-based) primitives. Use this resource when designing, reviewing, or debugging agent-driven applications to ensure reliable and context-efficient performance.

Diagnoses full-stack failure modes in agent and LLM applications, including wrapper regressions, memory contamination, tool discipline failures, hidden repair loops, and rendering corruption. Produces severity-ranked findings and code-first fixes for production agent systems.

A comprehensive guide covering patterns and architectures for automating code execution using LLMs. It spans simple sequential pipelines (e.g., daily dev scripts) to advanced, RFC-driven multi-agent Directed Acyclic Graph (DAG) orchestration. Ideal for building CI/CD style continuous development pipelines, managing complex state, and automating quality gates without human intervention.

A comprehensive pattern for building robust and cost-efficient LLM application pipelines. This guide details combining model routing based on task complexity, immutable cost tracking, limited retry logic, and prompt caching to manage API expenditures while maintaining high quality. Essential for production systems handling variable-complexity, high-volume tasks.

A high-performance utility for executing simple, deterministic code transformations (e.g., removing console logs, converting var to const). This engine leverages a WASM core to perform structural refactoring sub-millisecond, without calling any expensive LLM APIs. Ideal for CI/CD pipelines, automated bulk refactoring, and cost-sensitive development workflows where determinism and speed are paramount.

This tool routes tasks through a multi-tier system, specifically partitioning tasks into Agent Booster eligible (Tier 1) and standard LLM calls (Tier 2/3). It calculates and reports the estimated cost, leveraging the structural $0 cost of Booster bypasses for simple transforms, helping users optimize spend before running batches of similar tasks.

This skill analyzes historical benchmark run data stored in JSON files. It surfaces key performance drift metrics over time, including win rate, average latency, p99 latency, and LLM-baseline cost. It is crucial for detecting subtle performance degradations or regressions that standard smoke tests might miss, ensuring stability across releases or corpus expansions.

This guide details integrating FoundationModels for running Large Language Models (LLMs) directly on the device (iOS 26). It covers advanced patterns including text generation, extracting structured data using @Generable, implementing custom tool calling for domain logic, and streaming responses for real-time UI updates. Key features emphasize strong privacy and offline capability, ensuring no data leaves the device.

GateGuard is a mandatory pre-tool-use hook designed to significantly enhance the reliability and safety of LLM-generated code and actions. Instead of relying on self-evaluation, it forces the model to perform concrete, verifiable investigations—such as listing all imported files, checking data schemas, or detailing rollback procedures—before any write or destructive command execution. It establishes a three-stage gate (Deny, Force, Allow) to prevent hallucination and ensure context awareness in complex coding environments.

This skill automates the Test-Driven Development (TDD) repair loop. Given a failing unit test, it spawns a bounded, headless LLM agent to automatically modify the source code until the test passes. It closes the development loop by treating the test's pass/fail status as the definitive fitness function, ideal for CI failure triage and local TDD workflows.