Agentic Coding

Anthropic transferred the Agent Skills format to the Agentic AI Foundation in October 2026, making it a portable standard now supported by over 70 agent platforms including Cursor. This shift means SKILL.md files authored today can be shared across compatible agents, not locked to Cursor alone. The article provides a comprehensive 12-step, 45-minute walkthrough for building a skills library: creating three example skills (release-notes, pr-review-checklist, db-migration-safety), bundling executable scripts, writing reliable descriptions that trigger consistently, validating skills in CI/CD pipelines, and rolling out skills to teams with code-owner governance. Key practices include keeping skill descriptions under 1,024 characters with explicit "Use when" triggers, limiting skill bodies to under 5,000 tokens and moving reference material into separate files, treating skills as executable code requiring security review, and using scripts for deterministic work like parsing git history. The guide emphasizes that strong descriptions are the foundation—vague descriptions cause skills to fire unreliably or not at all—and shows how to test activation across single and multi-task prompts, troubleshoot common failures like mismatched folder names or personal skills that don't register in Cloud Agents, and maintain a lean catalog by retiring unused skills quarterly.

Gentle-AI is an open-source framework for agentic coding that implements Organic Driven Development (ODD) and Receipt-Driven Development (RDD) to optimize AI-assisted software development workflows. Key features include Engram for persistent project context across sessions, proportional task routing based on complexity rather than file count, and test-first development by default. The framework supports 17 AI agents including Claude Code and Cursor, with measured benchmarks showing significant improvements: a 3-line bug fix improved from 335 seconds and $0.339 to 77 seconds and $0.105 using inline work instead of subagents, while a data migration task achieved first-round acceptance with an independent verifier. ODD keeps small changes lightweight without planning overhead and delegates subagents only for specific reasons like exploration, parallel writing, context handoff, or independent verification—never by file count or task size. RDD provides proportionate review at the right depth with frozen candidates and deterministic workflows, while the Skills library and CodeGraph provide automated context loading and codebase symbol graphs respectively.