How Anthropic
Uses Claude Code

01

Give it a check to run

Claude stops when the work looks done. If it has nothing to run that says whether it actually is done, the docs put it plainly: you become the test suite.

So give it something that comes back pass or fail, like a test file, a build, a linter or a browser screenshot next to the mockup. Then Claude closes the loop itself. It makes the change, runs the check, reads the output and fixes it until it passes. And always ask for evidence, because "all tests pass" on its own tells you very little.

For long runs, /goal sets a condition a separate model checks after every turn, and Claude keeps going until it’s met.

02

Explore, then plan, then code

Letting Claude jump straight to code is how you get a beautiful solution to the wrong problem. Press Shift+Tab to cycle into plan mode, or start a message with /plan. In plan mode Claude reads files and explores, then writes a plan without editing your code until you approve it.

Put your effort into that plan. If something goes sideways mid-build, go back and re-plan instead of patching. My rule of thumb: if you can describe the change in one sentence, skip planning and just ask.

Tip inside the tip: press Ctrl+G to open the plan in your text editor and tighten it before Claude runs.

03

Turn mistakes into memory

Claude loads a file called CLAUDE.md into every session. It’s your project’s standing instructions. So when Claude gets something wrong, don’t just fix it this once. Ask it to add a rule.

If you just say "remember that", Claude saves it to its own auto memory. To put it in CLAUDE.md, say so directly, or edit the file with /memory. Keep the file lean. Anthropic recommends under 200 lines, and every line costs a little attention on every task.

04

Push it

Working isn’t the same as best. Treat the first version as a draft and push with these.

The docs call out be ambitious as a phrase that still helps, because Claude tends to build the smallest thing that satisfies your request.

05

Ship something tighter

After you’ve made changes, type /simplify. Four review agents run in parallel, checking reuse of existing code, simplification, efficiency and whether the change sits at the right level, then apply the fixes. (Commands reference)

/simplify doesn’t hunt for bugs. Its sibling /code-review does, in a fresh subagent that never saw your conversation. Add --fix to have Claude apply what it finds. One pass for quality, one for correctness.

06

Use subagents

For big jobs, ask for subagents. Each helper works in its own context and reports back only the answer, so your main conversation doesn’t fill up with everything it read.

Anthropic’s Growth Marketing team built a workflow with two specialized subagents, and it "generates hundreds of new ads in minutes instead of hours." One honest note from the docs. Subagents shine for parallel work and independent review, but each one costs extra usage, and handing the same task to a subagent doesn’t make the answer better.

07

Use /loop and /schedule

/loop reruns a prompt on an interval while your session stays open, like /loop 15m check CI and fix anything red. /schedule sets up a cloud scheduled task that keeps running with your laptop closed. It needs a claude.ai login, not an API key.

Anthropic’s Product Design team feeds Figma files to Claude Code and sets up autonomous loops where Claude writes the feature, runs the tests and keeps iterating, then they review. Watch the first few runs of anything new before you trust it alone.

08

Use /voice

Claude Code has voice dictation built in. Run /voice, then hold Space and talk. Your words show up in the prompt live, so you can say the big picture and type exact file names in the same message. Prefer not to hold? /voice tap starts and sends with a tap. (Voice docs)

It needs a claude.ai sign-in and a local microphone, and transcription doesn’t count toward your usage. In my experience, prompts I say out loud come out more detailed than ones I type, which is exactly what Claude needs.

01

Let Claude interview you

For anything big, don’t write the perfect prompt. Let Claude pull it out of you, then start a fresh session to build from the spec.

02

The writer and reviewer split

Have one session build it and a fresh one review it against your spec. A clean context isn’t swayed by the reasoning that produced the code.

03

Clear context between tasks

Most current models have a 1M-token window and compact on their own, so running out is rare now. /clear between unrelated tasks still gets you better answers and costs less. And if you’ve corrected the same thing twice, /clear and write one better prompt instead of a third correction.