Agentic Coding

This repository contains Claude Code's system prompts and related documentation, recently updated in January 2026 to include ~40 system reminders. The content details all of Claude Code's prompts, their token counts, and a comprehensive changelog tracking changes across 193 versions. It also documents the Piebald tool for advanced agentic development. For your specific interests, the repository provides system prompt documentation for advanced agentic workflows. Key relevant items include: Agent Prompt for agent creation architect (1110 tokens) for building custom AI agents, System Prompt for coordinator mode orchestration (3526 tokens) for delegating work to worker agents, and comprehensive memory instructions and consolidation prompts for maintaining persistent context. The repository also covers Plan mode (715 tokens) for enhanced planning subagents and detailed security monitoring prompts for autonomous agent actions. For PRD-based workflows, there are prompts for CLAUDE.md creation (384 tokens) analyzing codebases and agent-focused skill definitions. The documentation includes debugging guidance through the Debugging skill (417 tokens) and extensive tool descriptions for orchestrating complex agentic workflows, though it focuses on system prompt structure rather than best practices guides or video content.

The content covers best practices for creating effective Claude skills, which directly aligns with your interest in advanced work with AI agents. Key takeaways include structuring skills with a SKILL.md file containing metadata and instructions, using layered loading to keep only necessary information in context, and writing specific skill descriptions that tell Claude when to use them. The article emphasizes that most skills fail because they have vague descriptions or don't trigger reliably, not because they're technically complex. Critical practices include keeping skill bodies under five hundred lines, pushing detailed reference material into separate files linked one level deep, using the plan-validate-execute pattern for risky operations, and most importantly—building test evaluations before writing the skill itself. Anthropic's updated skill-creator tool now automates this evaluation loop with separate agents for execution and grading, allowing you to measure whether a skill actually improves performance against baseline scenarios. The piece also mentions security considerations when installing skills from untrusted sources, and introduces PM OS as a pre-built library of tested workflows and skills that eliminates the need to rebuild your agentic operating layer from scratch for common tasks like PRD reviews.

Claude Code memory files best practices, agentic coding PRD workflow setup, and AI pair programming debugging techniques are all well-represented in this curated repository. For memory files, see resources like `claude-md-templates` which provides CLAUDE.md best practices and project-specific configurations, and `claude-code-dna` which offers a behavioral OS with three-layer memory architecture inspired by mem0 and GraphRAG. For PRD workflow setup with agents, examine `mattpocock/skills` which includes PRD writing and `arun-mosai/claude-code-slice-skills` that transforms feature ideas into validated design docs through multi-agent orchestration. For debugging techniques in pair programming, `obra/superpowers` provides battle-tested debugging patterns and collaboration tools, while `sober-coding` offers a post-generation quality analyzer with 27 rules across seven dimensions targeting AI-native code smells. Additional resources include `voidborne-d/sober-coding` for code quality analysis and `disler/claude-code-hooks-mastery` for deterministic control and advanced debugging workflows. The repository emphasizes production-ready, actively maintained projects with clear documentation across all three areas of your interest.

The article describes six best practices for consistent Claude Code output across sessions. Key recommendations include creating a CLAUDE.md file before your second session to capture project rules and coding standards, keeping each session focused on a single bounded task to maintain output quality, and maintaining a MEMORY.md file to document decisions made during each session so knowledge persists across weeks. Additional practices emphasized are starting fresh with a clean session when context becomes messy rather than continuing to correct, providing specific output format descriptions instead of vague requests to minimize back-and-forth iterations, and organizing project documentation in a /docs folder so Claude Code can reference information independently. These practices compound over multiple sessions, with the setup reportedly enabling faster development than months of unstructured use.