Implement Agentic Engineering Best Practices with Claude for Coding

What this is, in plain English

“Vibe coding” is just chatting with an AI and hoping for good code; you get inconsistent results. “Agentic engineering” is the disciplined alternative: you write down, as files in your project, what your AI helpers are (subagents), what they can do (skills), and how to kick them off (commands). With that structure, Claude produces more consistent, reproducible code, because it follows defined roles instead of improvising each time.

Be honest about the level: this is an advanced practice for developers. You define and version these configuration files, and the source is a best-practices guide rather than a copy-paste tutorial (it leaves the exact prompts and execution to you). This page explains the idea, what it gives you, and the realistic shape of setting it up.

What you can use it for

• Consistent code across a project. Define once how functions, APIs, and errors should be written, and every AI result follows it.
• A maintainable MVP. Keep a solo-built codebase coherent as it grows, instead of a pile of one-off AI snippets.
• Team standardization. Everyone’s AI uses the same agreed roles and skills.
• Reproducibility. Because the setup is files in version control, you can review, share, and improve it like code.
• The reusable idea: writing down roles, skills, and commands turns ad-hoc AI use into a repeatable system.

Tools you need

• Claude (freemium): the AI you use to draft the definitions and do the coding. Free tier is enough to experiment.
• GitHub (freemium): version-control your code and your .claude definition files.
• Slack (freemium, optional): if you want to wire AI reviews or notifications into team chat.

How it actually works

You create a structure, fill it in with Claude’s help, then run your commands. The realistic shape:

1. Make the folders. In your project root, create a .claude folder containing three subfolders: agents, skills, and commands.

2. Define a subagent. A subagent is a specialist (say, a “Python function generator”). Have Claude draft its role and save it in .claude/agents/. A prompt to start from:

   Help me define a subagent for my coding project. It is responsible for generating
   Python functions. List its key responsibilities, inputs, and outputs, formatted as a
   markdown file for a .claude/agents/ directory.

3. Outline a skill. A skill is one concrete ability (like “code generation”). Have Claude draft it and save it as SKILL.md in a folder under .claude/skills/:

   Define a "code generation" skill that takes a function signature and description and
   produces the complete function. Describe the steps it follows, formatted as a markdown
   file for a .claude/skills/ directory.

4. Create an orchestration command. A command ties agents and skills together for a bigger goal. Save it in .claude/commands/:

   Define a command called "generate_feature" that uses my code-generation subagent and
   skill to turn a user story into working code. Describe the flow of information between
   them, as a markdown file for a .claude/commands/ directory.

5. Run your commands. With the definitions in place, invoke your command in Claude (via claude.ai/code or a CLI) to have it orchestrate the agents and skills you defined.

6. Review the code thoroughly. Whatever is produced, run a deep review before trusting it:

   Do an ultra-thorough code review of the following code: find bugs, suggest readability
   and performance improvements, and check it against best practices.
   [paste your code here]

Words you’ll see, explained

• Agentic engineering: structuring AI help with defined roles instead of ad-hoc chatting.
• Subagent: a specialist AI helper with one defined job.
• Skill: one concrete ability a subagent can use.
• Command: an instruction that coordinates subagents and skills toward a larger goal.
• Orchestrate: coordinate several pieces to run in the right order.
• Vibe coding: unstructured “just ask the AI and hope” coding, the thing this replaces.

Original source

Based on the claude-code-best-practice repository by shanraisshan on GitHub, which lays out moving from unstructured “vibe coding” to a systematic, agentic approach with defined agents, skills, and commands.

Notes & variations

• Do you even need this? For a quick one-off, plain prompting is fine. This structure pays off on real, growing projects where consistency matters.
• Free-tier viability: Claude’s free tier is enough to try these ideas, and GitHub is free for public repositories to host your .claude files.
• Common mistake: treating Claude as a plain chatbot for everything. Without defined roles, inputs, and outputs, results stay inconsistent.
• Tip for better results: start small, one agent and one skill for a narrow task, get it solid, then add commands. Keep all .claude definitions in version control, like code.