Agentic Coding

AAS Core is a local, agent-first control plane for AI coding assistants like Claude Code and Cursor that lets agents search and select from a complete catalog of 1,988+ skills without uploading your repository. The system validates agent-chosen skill stacks, generates reproducible aas-stack.json manifests, and produces immutable execution plans through a read-only MCP interface. Installation targets multiple platforms: use the AAS Core preview guide for Claude and Codex, or direct installers like `npx agentic-awesome-skills` for Cursor, Antigravity, Codex CLI, and others. Advanced workflows emphasize agent-owned composition where the AI inspects your project, compares skill candidates across capability areas (architecture, testing, security, deployment, UX), and creates evidence-backed selections validated before any changes. Specialized plugins package focused skill sets for domains like web apps, security, data analytics, DevOps, and QA. For efficient agentic coding, the system provides bundles (curated role-based skill groups), workflows (ordered execution playbooks), and deterministic planning—all designed to keep skill stacks reproducible, inspectable, and reviewable rather than one-shot opaque installations.

The dotnet-claude-kit is a free MIT-licensed Claude Code plugin designed for .NET 10 development that addresses four key Claude failures: legacy patterns, architecture roulette, token-expensive file reading, and missing quality gates. It ships with 47 skills teaching modern C# 14 patterns, 10 specialist agents, 16 slash commands, and a Roslyn MCP server with 20 compiler-query tools that reduce token costs by roughly 10x compared to file reading. Installation takes two commands, and /dotnet-init generates a project-specific CLAUDE.md that learns from your existing codebase rather than applying generic templates. The core daily workflow consists of /spec (agree on requirements), /plan (architecture-aware planning), /scaffold (complete features with 9-point checklist covering endpoints, validation, testing, OpenAPI), and /verify (7-phase quality gate: build, analyzers, anti-patterns, tests, security, formatting, diff review). The Roslyn MCP server is highlighted as the highest-value component, enabling semantic queries at 30-150 tokens versus 500-2,000 for file reads, making exploration on large codebases practical and token-efficient for agentic workflows.

Kimi Code offers agent-based coding workflows that go beyond traditional code suggestions, supporting project-level code understanding, multi-file editing, debugging, and seamless CLI and IDE integration—features particularly suited for complex software development projects. The article emphasizes expert practices for maximizing AI coding assistants, including providing clear and detailed prompts, sharing sufficient code context, reviewing AI-generated code before integration, and building consistent AI-assisted development workflows that integrate planning, coding, debugging, testing, and documentation. Other advanced tools mentioned include Augment Code with its Context Engine for analyzing codebase relationships and supporting agent-driven workflows, Cursor with its AI chat and multi-file editing capabilities, and Windsurf with its Cascade AI workflow for multi-step coding processes.

ECC (Extended Claude Code) is a comprehensive agent harness system for advanced agentic workflows with Claude Code, Cursor, and other AI coding tools. It provides 67 specialized agents, 279 reusable skills, 94 maintained commands, and production-ready hooks for token optimization, memory persistence, continuous learning, and verification loops. For your specific interests in efficient agentic coding with PRDs and AGENTS.md workflows, ECC offers: structured skill-based workflow surfaces with YAML frontmatter definitions; instinct-based continuous learning (v2) that auto-extracts patterns from sessions into reusable skills; session persistence hooks that save/load context automatically; memory optimization techniques including token capping and strategic compaction; and cross-harness support (Claude Code, Cursor, OpenCode, Codex, GitHub Copilot). The system is organized around agents for delegation, skills as the primary workflow surface, and rules as always-follow guidelines—all versioned at v2.0.0 (June 2026) with 261 skills and a control-pane substrate. Installation supports selective profiling (minimal, core, full) and works via plugin marketplace or manual install with language-specific rules copied to ~/.claude/rules/ecc/.