Wink Pings

Anthropic's 27-Page Guide: Claude Code Configurations from Top Teams—These 7 Are Most Worth Copying

Anthropic interviewed teams such as Higgsfield, ClickHouse, and Harvey, who shared the Claude Code configurations they use every day. Of the 27-page guide, these 7 tips are most worth copying—and each can be used directly.

Anthropic has a 27-page PDF about how top startup teams use its command-line AI coding tool, Claude Code, to ship products. There is no fluff in it—it contains the actual configurations that teams like Higgsfield, ClickHouse, and Harvey use every day.

Full PDF: Anthropic 27-page guide

Claude Code guide

This guide has a very high information density, and there is even a one-page checklist at the end. The 7 tips most worth copying are below:

1. Integrate tools instead of pasting
If you keep pasting content from some tool into Claude, connect the tool directly. If the tool itself has a good CLI (such as gh or psql), prefer having Claude use the CLI directly—it saves even more tokens than MCP (Model Context Protocol). Someone in the comments added: every time you paste, you're doing a step the agent could have done itself.

2. Break instructions into three layers
Put things that must never change in the root CLAUDE.md. Put conventions for a directory in the CLAUDE.md inside that folder. Put workflows used only occasionally in Skills. Don't stuff one-off workflows into the global configuration, and don't copy global principles into every folder.

3. Don't cut corners where hooks should be used
Anything that must happen every time should be handled by a hook. It will run no matter what the model decides. For example, block commits if tests haven't passed, or strip secrets before leaving the sandbox. You can't count on the model to remember these things every time.

4. Use /goal for long tasks
When a task is too long, Claude may declare completion too early or drift off course along the way. /goal forces alignment with the objective. It's suitable for large projects broken into multiple steps.

5. Learn to fan out multiple subagents
Have Claude split a large analysis across multiple parallel subagents, or send one subagent to decompose another agent's output. It's like code review, except the thing being reviewed is an agent.

6. Enter plan mode before rewriting
Press Shift+Tab to enter plan mode. Claude first reads the entire codebase, proposes an approach, and only then writes code. If you want to correct a wrong direction, this is the cheapest time to do it. If you wait until it's done and then change it, it's no longer a matter of a few dollars of tokens.

7. Use git worktree to rebuild projects
Build v2 in a new worktree and run it in parallel with v1; merge once the new one wins. You don't have to worry about production issues while making changes.

Some commenters asked which team in these configurations actually used git worktree; others asked how many subagents can run at once. The guide doesn't give specific numbers, and the exact configuration may differ from team to team. Someone also asked whether these 7 tips still hold when you're the only person left on the team.

发布时间: 2026-11-20 05:54