Agentic Coding

Use Your CLAUDE.md Properly by keeping it under 200 lines, leading with commands for testing, building, and linting rather than style rules, and treating it as standing instructions that earn their place on every turn. Plan Mode Is Not Optional on non-trivial tasks—use Shift+Tab twice to let Claude read the codebase and understand architecture before making changes, review the plan, and only then switch to coding mode. Context Is the Most Important Resource You're Managing: use /clear between distinct tasks frequently, use /compact with intent mid-task to preserve what matters, and recognize that long debugging sessions burn through context fast, causing the model to forget earlier conventions. Hooks Over CLAUDE.md for anything that must always happen deterministically, such as a PostToolUse hook in .claude/settings.json that fires your formatter after any edit. Build Skills for work you do more than once a week by saving reusable workflows in ~/.claude/skills/ (personal) or .claude/skills/ (project-specific), such as generating blog post drafts or creating PR descriptions. Use Subagents to keep contexts clean by running exploration and implementation in separate sessions to avoid failed approaches contaminating the context. Ask for Evidence, Not Confirmation by prompting Claude to return actual test output and command results rather than summaries. Register Your MCP Servers Early at the start of projects so Claude can access the whole system without manual intervention, and use detailed [Description] attributes on tool methods to ensure consistent tool invocation. Checkpoint Before Anything Destructive using Esc+Esc or /rewind before bulk operations like reindexing or schema migrations to enable clean rollbacks if needed.

Claude Code's context management requires monitoring token usage at 80% capacity and exiting sessions to restart for complex multi-file work. The 80/20 rule prevents context depletion failures like inconsistent code, repeated questions, and lost architectural decisions. Use /compact to compress conversation history after major features or at natural breakpoints, which works efficiently on Anthropic's first-party API through background session summaries. Your project-root CLAUDE.md file persists across sessions and should contain only essential context; path-scoped rules get summarized away during compaction. Divide work into context-sized chunks with handoff notes between sessions, delegate exploratory reading to subagents to preserve main session space, and use /context to monitor where tokens are spent. For long sessions, run context refreshes every 30 minutes and break tasks smaller rather than requesting larger context windows.

This guide teaches advanced Claude Code patterns, including context management strategies and agentic workflow design, which directly align with your interests. Key relevant sections: The guide covers context management thresholds (context pressure changes behavior at 70%, 85%, 90%), memory hierarchy through CLAUDE.md files, and agents vs skills vs commands decision frameworks. Section 9.20 covers Agent Teams coordination for parallel work on large codebases. The architecture guide explains the master loop and 4-layer context model. It includes 48 Mermaid diagrams visualizing multi-agent topologies and memory management patterns. The methodologies section (TDD/SDD/BDD) addresses maintaining code quality with AI assistance. For enterprise adoption, it covers team deployment strategies, role-specific learning paths (Tech Lead, CTO, Engineering Manager), and systematic MCP vetting workflows with 28 CVE tracking and 655 malicious skills database. The guide includes 181 annotated templates (23 agents, 64 skills, 37 hooks) with explanations of why patterns work, plus a 271-question quiz for knowledge validation across 9 categories.

This repository contains Claude Code's system prompts extracted from the latest npm version. As of Claude Code v2.1.201 (July 3rd, 2026), it lists 515+ system prompts with token counts, a changelog tracking 227 versions, and is updated within minutes of each Claude Code release. The repo highlights that Claude Code uses multiple conditional system prompts rather than a single string—including descriptions for builtin tools (Write, Bash, TodoWrite), separate prompts for subagents (Explore, Plan), and numerous AI-powered utility functions. For users wanting to customize prompts, the repository recommends tweakcc, which lets you modify individual prompt pieces as markdown files and patch your Claude Code installation while managing conflicts with Anthropic updates.