claude-code-best-practice

practice makes claude perfect

CONCEPTS

Note: Custom slash commands have been merged into skills. Files in .claude/commands/ still work, but skills (.claude/skills/) are recommended as they support additional features like supporting files, invocation control, and subagent execution.

• Skills - Reusable knowledge, workflows, and slash commands that Claude can load on-demand or you invoke with /skill-name
• Subagents - Isolated execution contexts that run their own loops and return summarized results
• Memory - Persistent context via CLAUDE.md files and @path imports that Claude sees every session
• Rules - Modular topic-specific instructions in .claude/rules/*.md with optional path-scoping via frontmatter
• Hooks - Deterministic scripts that run outside the agentic loop on specific events
• MCP Servers - Model Context Protocol connections to external tools, databases, and APIs
• Plugins - Distributable packages that bundle skills, subagents, hooks, and MCP servers
• Marketplaces - Host and discover plugin collections
• Settings - Hierarchical configuration system for Claude Code behavior
• Permissions - Fine-grained access control for tools and operations

Extension Overview: See Extend Claude Code for when to use each feature and how they layer together.

MY EXPERIENCE

■ Workflows

• Claude.md should not exceed 150+ lines. (still not 100% guarenteed)
• use commands for your workflows instead of agents
• have feature specific subagents (extra context) with skills (progressive disclosure) instead of general qa, backend engineer.
• /memory, /rules, constitution.md does not guarentee anything
• do manual /compact at max 50%
• always start with plan mode
• subtasks should be so small that it can be completed in less than 50% context
• vanilla cc is better than any workflows with smaller tasks
• commit often, as soon as task is completed, commit.

■ Utilities

• iTerm terminal instead of IDE (crash issue)
• Wispr Flow for voice prompting (10x productivity)
• claude-code-voice-hooks for claude feedback
• status line for context awareness and fast compacting
• git worktress for parallel development
• /config dont ask permission mode instead of dangerously—skip--permissions

■ Debugging

• /doctor
• always ask claude to run the terminal (you want to see logs of) as a background task for better debugging
• use mcp (claude in chrome, playwright, chrome dev tool) to let claude see chrome console logs on its own
• provide screenshots of the issue

CONTEXT ENGINEERING

• Humanlayer - Writing a good Claude.md
• Claude.md for larger monorepos - Boris Cherny on X

WORKFLOWS

• Ralph plugin with sandbox
• Human Layer RPI - Research Plan Implement
• AgentOs - 2026 its overkill (Brian Casel)
• Github Speckit
• GSD - Get Shit Done
• OpenSpec OPSX
• Superpower
• Andrej Karpathy Workflow
• Creator of Clawd Bot Workflow

KEYWORDS

Claude: btw (background task) • "defaultMode": "bypassPermissions" • ultrathink (deprecated)

Community: Agentic Workflow • AI Slop • Context Engineering • Context Rot • Dumb Zone • Hallucination • One Shot • Rate Limit Jail • Slot Machine Method (save→run→revert→retry) • The Holy Trinity (Skills+Agents+Hooks) • Token Burn • Vibe Coding

CLAUDE CODE FEATURES INSPIRATION

• Claude Code Tasks - inspired by beats Inspiration
• Ralph Plugin

COMMAND + SKILL + SUBAGENT ARCHITECTURE

Component	Role	Example
Command	Entry point, user interaction	/weather-orchestrator
Agent	Orchestrates workflow with preloaded skills	weather agent
Skills	Domain knowledge injected at startup	weather-fetcher, weather-transformer

When to use: Multi-step workflows • Domain-specific knowledge injection • Sequential tasks • Reusable components

Why it works: Progressive disclosure • Single execution context • Clean separation • Reusability

See weather-orchestration-architecture for implementation details.

SETTINGS

See reports/claude-settings.md for a comprehensive reference of all available settings.json configuration options.