Agentic Coding

This course teaches operator-level agentic coding for developers who already use AI agents casually. It covers reading and understanding agent sessions including tool definitions and rule files, optimizing context windows to show models only what they need, maintaining unified agent rules across different harnesses, writing skills that load conditionally, automating mechanical work like formatting and linting through hooks, consolidating external services into a single tool layer, and running multi-phase workflows end-to-end including research, planning, parallel implementation, review, QA and pull requests. The lessons are free, build progressively on context models from previous lessons, and primarily use Claude Code examples with notes on Anthropic-specific features.

Claude Code agentic workflows transform AI assistance from simple code generation into comprehensive repository-level engineering. Key practices include separating exploration from execution—first understanding the codebase before making changes—and using explicit planning phases where Claude Code proposes an implementation strategy for developer review before any code modification. The workflow emphasizes the "Explore → Plan → Implement → Test → Debug → Review → Approve" pattern, treating testing as integrated feedback rather than a final step, with Claude Code analyzing test failures and iterating toward working solutions. For context management, focus on relevant subsystems rather than dumping entire repositories, and document stable engineering rules as reusable project guidance to reduce repeated instructions. Productivity gains come from delegating coordination overhead—repository searching, test generation, failure investigation, documentation—rather than pure code generation speed, allowing engineers to focus on architecture, risk assessment, and system design while Claude Code handles execution layer tasks.

# Claude Code System Prompts Repository This GitHub repository documents Claude Code's complete system prompt architecture, expanding from 350 to 515+ prompts across versions. The repository tracks all Claude Code system prompt changes automatically and offers the tweakcc tool for customizing individual prompt pieces via markdown files before patching npm or native installations. The resource covers agent prompts for subagents (Explore, Plan), slash commands (/code-review, /batch, /schedule), creation assistants (CLAUDE.md generation, status-line setup), and utility functions. It also documents builtin tool descriptions, data references (API SDKs, Managed Agents, gateway protocols), system reminders, and 80+ skills for specialized tasks including code review, design sync, artifact creation, and debugging. For context management, the repository demonstrates how Claude Code structures conditional prompt sections based on environment configs, tool availability, and session state—providing a concrete reference for efficient agent prompt organization. The tweakcc tool enables custom modifications while maintaining sync with official updates.