Building With Claude Code: A Practitioner's Field Guide
A practitioner's guide to building real software with Claude Code: the working habits, the agent architecture (tool vs skill vs subagent), and the prompting discipline that make it reliable.
Guillaume Rufenacht
AI Product Manager · Lisbon
Claude Code isn’t autocomplete for code. It’s an agent with access to a computer, and used well, it changes the unit of work from “type faster” to “direct a team of agents.” This is my practitioner’s guide to building with it: the habits, the agent architecture, and the prompting, each linking to a deeper piece.
I build real products with Claude Code, from the AI outbound system to the LLM pipelines behind Geonimo. Here’s the map of how to get the most out of it.
Key takeaways
- Treat Claude Code as 'Claude with a computer,' not a code-completion box.
- Fix the system (CLAUDE.md, skills) instead of re-prompting; make verification agent-native.
- Choose the right primitive as agents grow: tool, skill, or subagent.
- Prompt like an engineer: start from evals, give tools, state trade-offs.
The habits: how to actually work with it
The biggest gains aren’t prompt tricks, they’re working habits: fix the system instead of correcting one response, make verification something the agent can do itself, trust good defaults so you can run several agents at once, and keep context lean. These come straight from how Anthropic’s own team works, broken down in how Anthropic’s own team uses Claude Code.
The architecture: tool, skill, or subagent
Agents rot when you keep bolting capability onto a growing system prompt. The fix is reaching for the right primitive: human-like primitives first (file system, code execution), then skills for progressive disclosure, then subagents only for parallelism or a fresh perspective. The decision framework is in tool, skill, or subagent?.
The prompting: like an engineer, not a pleader
Good prompting for Claude Code is an engineering discipline: measure with evals, clean up structure before optimizing, give the model tools instead of pleading with it, and state both sides of any trade-off. The full set of techniques is in the Claude Code prompting playbook.
Putting it together
These reinforce each other. Good prompts and the right primitives keep your agent simple; agent-native verification lets it run unattended; fixing the system means it improves every time instead of needing you every time. The AI SDR system is one concrete example of all of it working together.
The takeaway
This is how I ship with Claude Code in production. See the stack I build with or get in touch if you want help building this way.
Frequently asked questions
What is Claude Code?
Anthropic's agentic coding tool, essentially Claude with access to a computer (file system, code execution, web search). It's used to build and ship real software, not just autocomplete code.
How do I get good at Claude Code?
Fix the system instead of re-prompting (use CLAUDE.md and skills), make verification agent-native, prompt from evals, and choose the right primitive (tool, skill, or subagent) as your agent grows.
Do I need to be an engineer to use it well?
Less than you'd think. The leverage is in product taste, clear problem framing, and verification; the agent writes the code.
Work with me
Want a system like this built for your pipeline?
I help teams take AI from a clever prototype to dependable production, outbound engines, lead intelligence, and the LLM pipelines underneath. See what I have shipped or get in touch.