Claude Code Best Practice: agentic engineering guide
Claude Code Best Practice is a curated reference for using Anthropic's Claude Code coding assistant well, covering concepts, workflows, and 83 tips in one GitHub repo with 64,049 stars. Reach for it once you're already running Claude Code daily and want a checklist of concrete tactics instead of a first tutorial; skip it if you need runnable code or a guided course, because it's pointers and prose, not software.
What is Claude Code Best Practice?
Claude Code Best Practice is a GitHub repository maintained by shanraisshan that catalogs concepts, workflows, and tips for Claude Code, Anthropic's terminal-based coding agent. Its README works like an index: tables map each Claude Code concept (subagents, commands, skills, hooks, MCP servers) to its file location and official docs, then list 83 tips plus links to a dozen related workflow repos.
Key Areas of Guidance
- ✓A Concepts table maps every core Claude Code building block — subagents, commands, skills, hooks, MCP servers, plugins, settings, memory, checkpointing, and CLI flags — to its file location and official docs page, with a 'Hot' subsection tracking newer additions like Ultrareview, Auto Mode, the Agent SDK, and Git Worktrees.
- ✓A Development Workflows section built around one shared pattern (Research → Plan → Execute → Review → Ship), comparing a dozen other developer workflows in one table, from Superpowers down to HumanLayer.
- ✓A Cross-Model Workflows section spelling out three concrete mechanisms for LLM orchestration — plugin, MCP, and router — for running Claude Code alongside Codex, Gemini, or other models.
- ✓Skill Collections and Agent Collections tables pointing to other curated AI agent skills and subagent libraries, sorted by star count, for when the built-in examples aren't enough.
- ✓83 tips and tricks split into categories like Prompting, Planning/Specs, Context, and Session Management, each tagged 🚫👶 where the author trusts Claude to run a step unsupervised.
- ✓Named practitioner writeups instead of anonymous crowd-sourced advice: dated tip sets attributed to Boris Cherny, Claude Code's creator, plus workflows credited to Andrej Karpathy and Peter Steinberger.
- ✓The Concepts table's Best Practices row links straight to Anthropic's own prompt engineering tutorial rather than re-explaining it.
Who is This Resource For?
Developers who already use Claude Code regularly and want a concrete checklist — CLAUDE.md structure, context thresholds, when to spin up a subagent — get the most out of Claude Code Best Practice. Teams standardizing on one agentic workflow across a codebase can use the Development Workflows comparison to pick a starting pattern. Skip it if you've never run `claude` before: the repo assumes working knowledge of commands, hooks, and skills rather than teaching them from zero.
Strengths
- ✓One table maps every core Claude Code concept (subagents, commands, skills, hooks, MCP, plugins, settings, memory) straight to its file path and docs link, which beats hunting through code.claude.com yourself.
- ✓83 tips are grouped by real workflow stage, not dumped in one list — Context, Session Management, and CLAUDE.md rules are separate categories you can jump to.
- ✓The Cross-Model Workflows section names the actual mechanism (plugin vs MCP vs router) instead of a vague 'use other models too.'
- ✓Tip attribution is specific: dated writeups tied to Boris Cherny, Claude Code's creator, rather than unsourced 'best practices.'
- ✓MIT licensed, so you can fork or copy sections into your own team wiki without a licensing conversation.
What This Repository Isn't
- △It's a link-heavy index, not standalone teaching content — most Concepts rows point out to code.claude.com docs or other repos rather than explaining the mechanism in the file itself.
- △The Development Workflows, Skill Collections, and Agent Collections tables are curated by one maintainer's judgment call on what's worth listing, with no visible changelog for when tips get added, revised, or dropped.
- △Nothing here is runnable: there's no install step or package to add to a project, so it doesn't help if you want executable examples rather than reading material.
- △GitHub lists HTML as the repo's primary language rather than Markdown, per the repo facts — a signal that some content renders through HTML tables and badges rather than plain prose, which can make raw diffs harder to skim.
Related AI Coding Resources
Frequently Asked Questions
In Claude Code Best Practice's own framing, vibe coding means prompting an AI assistant loosely and accepting whatever it produces, without a plan or review step. The repo's description sets that against 'agentic engineering' — deliberate practices like plan mode, phased specs, and review loops — and positions itself as the path from one to the other.
Claude Code Best Practice centers on Claude Code, but its Cross-Model Workflows section documents three ways to bring in other models — a plugin (another CLI's slash commands inside Claude Code), an MCP server, or a router that swaps the API endpoint — with named example repos for pairing Claude Code with Codex, Gemini, or other providers.
The Concepts section of Claude Code Best Practice is a table, not prose: each row names a building block — subagents, commands, skills, hooks, MCP servers, plugins, settings, memory, checkpointing, CLI flags — next to its config file location and a link to the official Claude Code docs, plus a 'Hot' subsection for newer features like Auto Mode and Git Worktrees.
Claude Code Best Practice is released under the MIT license, per its GitHub repository metadata.
Claude Code Best Practice has a dedicated Context tips category: it notes that, on the 1M-context model, context rot is observed to set in at approximately 300-400k tokens, describes a 'dumb zone' near 40% context usage where output quality degrades, and recommends resetting with /compact or /clear rather than letting autocompact fire mid-task.
Best use cases
- •Onboarding a new hire onto Claude Code — hand them the Concepts table instead of five separate doc links.
- •Picking a workflow methodology before you commit engineering time — compare a dozen approaches in the Development Workflows table rather than trying each one.
- •Debugging degrading output on a long session — jump straight to the Context tips category for the ~300-400k token context-rot threshold and the ~40% 'dumb zone' guidance.
- •Wiring Claude Code to review its own work with a second model — the Cross-Model Workflows section gives you the plugin/MCP/router options and named repos to start from.
- •Writing or trimming your own CLAUDE.md — the CLAUDE.md + rules tip category has specific line-count guidance (under 200 lines, one example at 60) instead of generic advice.