Claude Code
The CLI agent from Anthropic that walks alongside me every day. This article is a usage report — credit for the tool goes entirely to Anthropic.
What it is
Claude Code is a command-line agent that runs in the terminal and can read, write, execute code and shell commands in the working directory. It drives the Claude model (Opus, Sonnet, Haiku) with a system of tools, skills and sub-agents.
From the user's side, you talk to it like a pair programmer: you describe the goal, it reads the code, it proposes a plan, it edits the files, it runs the tests. It's integrated with VSCode and JetBrains via an extension, which keeps the IDE context.
How it works
Everything rests on a simple idea: the model only sees its context. Too much context creates noise — the model then has to filter before reasoning. Optimising means giving it less text but more relevant text.
You don't save tokens by asking 'reply shorter'. You save them by building a workflow that avoids the wrong context.
— My internal memo
Concretely, Claude Code loads into its context: the system prompt (stable), the project's and user's CLAUDE.md (stable if well-managed), the conversation history, and the results of invoked tools (Read, Grep, Bash, Edit, etc.).
The stable part benefits from prompt caching on Anthropic's side: with a frozen system prompt and CLAUDE.md, cached input tokens cost roughly 0.1× the standard price (cache write at 1.25× or 2× depending on TTL). On a project I work on all day, that's an input-side saving of around 90%.
How to install it
Official installation via npm, then authentication:
npm install -g @anthropic-ai/claude-code
claude # first run → login
The "Claude Code" VSCode extension lives in the marketplace, it auto-connects to the already-authenticated CLI. For project conventions, I put all rules in a CLAUDE.md at the root — short, readable, not a dumping ground.
How I use it daily
My usage always follows the same sequence, regardless of the task: pick a level (direct / lean / standard / full), load the right context, explore before editing, delegate when useful, measure.
- Tiny task (rename, oneshot): direct edit, no plan, no sub-agent.
- Normal feature: a quick brainstorm, then a plan I execute.
- Sensitive topic (auth, payment, customer data, AI, pricing): a full /call-tech-lead with independent sub-agents and guardrails.
- Always: read before writing, run tests before committing.
The combo that works well for me: Claude Code orchestrates, Superpowers provides discipline (skills + sub-agents), RTK trims command output, and monitor-ccu measures cost in near real time. This very site was rebuilt with this approach — Plans 1 to 4.
Sources & credits
Claude Code is designed and maintained by Anthropic. This article is purely a usage report — full credit for the product goes to Anthropic and the Claude Code team.