How Top Engineers Use Claude Code (Claude CLI)

Anthropic's terminal-based AI coding agent, Claude Code (Claude CLI), is changing how engineers around the world build software. In this article we review how leading developers actually use it, based on a survey of English-language blogs.

This article summarizes, in English, the key points we gathered from English-language technical blogs and official documentation. The source for each view is listed in the "References" section at the end of the article. Please always consult the original articles for full details.

1. Armin Ronacher (creator of Flask) — Write "code that LLMs can read"

Armin Ronacher, the creator of Flask, runs Claude Code on a monthly Max plan and deliberately makes the lightweight Sonnet model his primary choice rather than the more powerful Opus. It reflects a focus on balancing speed with sufficient quality.

What is especially distinctive is that he prioritizes having the agent write "readable" code in the first place. On the backend he prefers explicit, simple Go, favors plain SQL over an ORM, simple structures that avoid inheritance, and somewhat longer function names. These are deliberate choices so that the agent can correctly understand the codebase. He also has the agent use shell scripts and Makefiles rather than MCP, and always writes output to log files so that the agent can diagnose itself (source 1).

→ Read the dedicated article: How Flask creator Armin Ronacher uses Claude Code [with concrete examples]

2. Simon Willison (co-creator of Django) — Design the "agentic loop"

Simon Willison, co-creator of Django, cites "designing agentic loops" as an essential skill for engineers going forward. The idea is that an agent is "something that runs tools in a loop to achieve a goal," and that by choosing the right tools and goal you can draw out solutions through trial and error.

To that end, he emphasizes giving the agent verification methods it can run by itself. He notes that with a robust test suite, the agent can autonomously debug failing tests and iterate. At the same time, he recommends the high-speed mode that skips permission prompts only on the conditions of an "environment isolated with something like Docker" and "strict limits on any billable credentials" (source 2).

→ Read the dedicated article: How Django co-creator Simon Willison uses Claude Code [with concrete examples]

3. Boris Cherny (creator of Claude Code) — Carve mistakes into CLAUDE.md

You can learn how Boris Cherny, a creator of Claude Code itself, works from a community site that aggregates his public posts. He runs several Claude instances in parallel and describes a team habit of "every time Claude does something wrong, adding that to CLAUDE.md so it isn't repeated next time."

• Starts many sessions in plan mode and switches to automatic execution once the plan is settled
• Places custom slash commands such as /commit-push-pr in .claude/commands/ and shares them across the team
• Emphasizes that the setup is "surprisingly vanilla" and that Claude Code works well as-is by default
• Says that giving Claude verification methods "doubles or triples the final quality" (source 5)

In another interview, he mentions that with newer models explicit planning is needed less often, so he is moving from plan mode to "auto mode," and that he has shifted from multiple tabs to isolated sessions using git worktrees (source 6).

→ Read the dedicated article: How Claude Code creator Boris Cherny himself uses it [with concrete examples]

4. Mitchell Hashimoto (co-founder of HashiCorp) — Don't leave the agent idle

Mitchell Hashimoto, co-founder of HashiCorp and creator of the terminal emulator Ghostty, described his principle in an interview with The Pragmatic Engineer: "If I'm writing code, I want the agent to plan. If the agent is writing, I want to review." The idea is to structure time so that neither the human nor the agent ever sits idle.

He also emphasizes delegating research—investigation, edge-case analysis, library comparisons—and not just code generation, and advises engineers who are skeptical of AI to "start with research assistance rather than code generation" (source 4).

→ Read the dedicated article: How Ghostty creator Mitchell Hashimoto uses AI agents [with concrete examples]

5. Best practices recommended by Anthropic

These individual practices line up with the recommendations in Anthropic's official documentation. What the docs repeatedly stress is the fact that "the context window is the biggest constraint."

• Use /clear between unrelated tasks and /compact during long sessions to tidy up context
• Delegate codebase exploration to sub-agents and return only summaries to the main thread
• Work through the four phases of Explore → Plan → Code → Commit
• Keep CLAUDE.md short ("remove any line that wouldn't lead to a mistake if deleted")
• Always provide verification that Claude can run on its own, such as tests, builds, linters, and screenshot comparisons (source 3)

Common themes

The common points that emerge from the multiple sources we surveyed can be organized as follows.

• Context is a scarce resource — keep CLAUDE.md short and push exploration to sub-agents
• Give the agent a verification loop it can run itself — this is the single most important factor determining quality
• Use CLAUDE.md as "accumulating memory" — turn each mistake into a rule and share it via git
• Pair permission-skipping with an isolated environment — Docker and a spending cap are prerequisites
• A delegating mindset — entrust the work as you would to an engineer, rather than dictating line by line