Agentic Coding

The article presents a practical solution to a core problem in long-running coding agent workflows: maintaining intent continuity across extended project histories without relying solely on longer context windows. The author demonstrates that coding agents often fail to apply constraints stated in earlier conversations because they lack a mechanism to verify which past requirements still apply to new tasks—not because of memory limits, but because they can't automatically determine relevance. The proposed system uses a pure Python pipeline with three key components: extracting structured requirements from chat history, retrieving candidates based on a fixed domain schema that maps component relationships, and crucially, a verification layer that identifies superseded rules before passing context to the agent. Testing across eight tasks requiring knowledge of twelve planted requirements showed the baseline approach passing zero tasks, naive keyword search passing four, and the intent-aware verification method passing all eight—achieving 100% recall while using only 28% more tokens than naive retrieval. The implementation deliberately avoids embeddings, vector databases, and external LLM calls to isolate the verification mechanism's effectiveness, making it directly reproducible and applicable to advanced agentic coding workflows where project rules persist across sessions.

Agentic Awesome Skills is a repository and control system for managing reusable agent skills and prompts for AI coding assistants like Claude Code, Cursor, and others. The core product, AAS Core (currently in agent-first preview as of v17.5.0), enables agents to search and select from a catalog of 2,126+ skills, with the system validating and recording those choices in reproducible stack manifests (aas-stack.json) before execution. The workflow emphasizes agent-owned selection from the complete local catalog without uploading repositories, followed by human review of the validated stack and immutable plan before installation. For enterprise adoption and context management, AAS Core uses a local MCP (Model Context Protocol) interface with read-only compose_stack validation, allowing agents to inspect projects and choose exact skill IDs while keeping selections reviewable and deterministic. The system provides specialized plugins for focused domains (web apps, security, data analytics, DevOps, QA, SaaS launch), direct skill installation for multiple platforms (Claude Code, Cursor, Antigravity, Codex, Gemini CLI, GitHub Copilot), and management tools like the CLI validator and browser-local Workbench for reviewing artifacts without executing installations. For solo SaaS builders and teams, the repository offers role-based bundles, execution workflows for outcomes like shipping MVPs or security audits, and explicit guidance on reducing context window load through selective skill activation—addressing both efficient agentic coding workflows and practical enterprise constraints around token budgeting and deployment safety.