Autonomous loop patterns for multi-step AI workflows without human intervention. Use when building CI-style pipelines, parallel agent coordination, or continuous autonomous development cycles. Covers 5 loop architectures (Sequential Pipeline, Infinite Agentic Loop, Continuous PR Loop, De-Sloppify Pass, RFC-Driven DAG) with decision matrix.
Karpathy-style automated self-improvement loop for the .agent/ skill system. Use to measure skill quality, identify weaknesses through structured evals, generate targeted improvements, and verify them. The AI equivalent of gradient descent applied to agent instructions — works through Measure -> Find Weaknesses -> Generate Improvements -> Score -> Regression Check -> Iterate.
Designs server-side architecture, API contracts, and data models for production-grade services. Use for backend architecture decisions, service/repository implementation, and designing REST/GraphQL API contracts. Differentiator: layered architecture with documented ADRs, machine-readable API contracts (OpenAPI/GraphQL), and database patterns with connection management, transactions, and migration strategies.
Structured exploration before committing to an approach. Use when facing ambiguous problems, new features, architectural decisions, or any situation where multiple approaches are possible. Differentiator: Mandatory search-first integration before any idea generation, and a hard user-confirmation gate that prevents premature implementation.
Designs and configures CI/CD pipelines for GitHub Actions and GitLab CI with caching, matrix builds, security scanning, and deployment gating. Use for any continuous integration or deployment pipeline task — from initial setup through production release gating. Distinguishes itself through hardened pipeline patterns with cache optimization, secret management, flaky test detection, and staging-gated deployment.
Refactors and improves code quality with measurable before/after improvements in readability, complexity, and duplication. Use for cleaning up and optimising existing code, reducing cyclomatic complexity, and extracting maintainable patterns. Differentiator: behaviour-preserving refactoring with characterisation tests, before/after code examples for 7+ common code smells, and structured verification gates.
SHA-256 content hash caching for file processing to avoid redundant work and reduce LLM costs. Use when processing files (extraction, transformation) to avoid redundant work and reduce costs. Covers content-hash-based cache key design, frozen dataclass entries, file-based cache storage with O(1) lookup, and service layer wrapper pattern.
Context window preservation and session continuity skill for AI coding agents. Use proactively whenever commands, files, or API responses would produce more than ~20 lines of output, and reactively when context is nearing exhaustion or a session has been compacted. Orchestrates the context-mode MCP plugin (ctx_execute, ctx_index, ctx_search, ctx_batch_execute, ctx_execute_file, ctx_fetch_and_index) to keep raw data out of the LLM context window and retrieve only what's needed on demand.