The 83 tips that reveal how AI agents actually work

The most useful manual for Claude Code is not from Anthropic. It is a GitHub repo called shanraisshan/claude-code-best-practice, maintained by a community of power users. It collects 83 tips, a taxonomy of agent architecture, and a growing library of workflows, skills, and subagents. The repo’s subtitle is “from vibe coding to agentic engineering.” That framing is the story.

Vibe coding is the phrase Andrej Karpathy coined in early 2025 for the experience of typing a vague prompt and watching an AI generate code you barely understand. It works for prototypes. It fails for production. The repo documents what comes next: the deliberate, structured practice of building with AI agents as a craft.

The repo catalogs six core concepts that form the architecture of a Claude Code project: subagents, commands, skills, workflows, hooks, and MCP servers. Each maps to a file or directory convention. Subagents live in .claude/agents/<name>.md. Commands live in .claude/commands/<name>.md. Skills live in .claude/skills/<name>/SKILL.md. These are not abstract categories. They are files checked into git, versioned, and shared.

The distinction between skills and commands is instructive. A command is a slash-triggered workflow you run many times a day. A skill is a folder with a SKILL.md file, optionally containing references/, scripts/, and examples/ subdirectories for progressive disclosure. The repo advises: “if you do something more than once a day, turn it into a skill or command.” That is the operational definition of agentic engineering.

The tips section contains the most actionable material. Tip 1 on planning: “always start with plan mode.” Tip 2: “start with a minimal spec or prompt and ask Claude to interview you using AskUserQuestion tool, then make a new session to execute the spec.” This is the opposite of vibe coding. It is structured decomposition.

Tip 7 on planning is worth quoting: “break PRDs into vertical slices (tracer bullets) that cross all layers (DB + service + UI) — AI defaults to horizontal phasing (DB phase, then API phase, then frontend phase) which delays end-to-end feedback until the last phase.” This is a concrete observation about model behavior. The model defaults to the wrong decomposition strategy. The user must override it.

The context management tips are the most technically specific. Tip 1: “context rot kicks in around ~300-400k tokens on the 1M context model.” Tip 2: “dumb zone kicks in around ~40% context.” Tip 3: “aggressively keep it below 30%.” These are numbers with practical force. A user who ignores them will get degraded output and blame the model. A user who respects them will get consistent results.

The session management tips describe a decision tree. “Every turn is a branching point — after Claude ends a turn, pick between Continue, /rewind, /clear, /compact, or Subagent.” The repo treats each session as a state machine. The user’s job is to choose the right transition.

The CLAUDE.md tips reveal a tension. Tip 1: “CLAUDE.md should target under 200 lines per file.” Tip 2: use .claude/rules/*.md for auto-loading. Tip 3: “wrap domain-specific CLAUDE.md rules in <important if="..."> tags.” The repo is documenting workarounds for a model that ignores instructions when files grow long. That is a real constraint on how teams can use the tool.

The cross-model workflow section is the most forward-looking. Three mechanisms exist: plugin, MCP, and router. The plugin approach runs another model’s CLI inside Claude Code via slash commands. The MCP approach calls another model as a tool through the Model Context Protocol. The router approach swaps Claude Code’s API endpoint to a different provider. The repo links to musistudio/claude-code-router (34k stars) and openai/codex-plugin-cc (18k stars). These projects are building the infrastructure for multi-model agent systems.

The repo lists 16 major workflow repositories ranked by stars. The top entry is “Superpowers” at 236k stars. “Everything Claude Code” is at 220k. “Matt Pocock Skills” is at 142k. These numbers signal something: the community is investing heavily in structured methodologies, not ad-hoc prompting.

The repo also catalogs 10 skill collections and 2 agent collections. The largest skill collection is anthropics/skills at 154k stars. The largest agent collection is msitarzewski/agency-agents at 115k stars. These are not toy projects. They are production-grade libraries of reusable agent definitions.

What the repo does not say is as important as what it does. It does not claim that Claude Code works out of the box. It does not claim that vibe coding scales. It does not claim that the model understands your codebase without help. Every tip is a compensation for a model limitation. Every workflow is a defense against context rot. Every skill is a way to encode knowledge the model does not have.

The practical conclusion for AI builders is this: the gap between a demo and a production agent is a set of conventions. The conventions are not secret. They are documented in a GitHub repo with 83 tips and a growing taxonomy. The work of agentic engineering is learning to use them.

The repo’s final section lists 14 workflow methodologies, each with a star count and a step diagram. The most starred workflow, Superpowers, has 236k stars and 14 sub-loops. The least starred, HumanLayer, has 11k stars and 27 sub-loops. The variance in complexity suggests the field has not converged. The repo is a snapshot of a community figuring out what works.

That is the real story. Not that Claude Code is powerful. That the community has decided the power only comes with structure.