Anthropic just revealed a 24-minute master class on how to actually tip Claude

Claude Code: A Practical Guide from the Creators

At the Anthropic conference, Boris - one of the creators of Claude Code - held a practical session and told how to get started, what products to use and how to build a workflow. Publish key ideas without water. *

## What is Claude Code

Claude Code is not just another AI assistant with hints in an IDE. It’s a completely agent-driven tool: it doesn’t offer the next line of code, but instead fixes bugs entirely, plans and executes tasks from start to finish.

There is no need to change the usual workflow. Claude Code works in the terminal and integrates with VS Code, Xcode, JetBrains, locally, via SSH, through Tmux - in any environment.

## Where to start: Q&A codebase

If you’ve just opened Claude Code and don’t know what to type into a prom bar, start with questions about your code.

This is the #1 recommendation from the Anthropic team. This is how they train new technicians: From day one, they ask questions about the codebase through Claude Code. Previously, onboarding took 2-3 weeks. It's 2-3 days now.

What can I ask

• How is this class/function used?
• How do you establish this dependence?
• Why this argument was added – look at the git story.
• What did I do this week? (Claude looks at the git log and knows your username)

Claude Code does not do text search. It goes deep: looking for instances of instantialization, looking through git history, raising related issues on GitHub via WebFetch. Answers are obtained at the documentation level, not Cmd+F.

**Importantly, there is no indexing. Your code is not downloaded anywhere, is not used to train models, is not stored in databases. He's staying with you.

## How to properly compose smears

### Ask to plan before writing code

The most common mistake is to immediately ask to implement a big feature. Claude can cope, but the outcome is unpredictable.

It works better

Before you write code, make a plan. Show it to me and wait for confirmation.

That's enough. No need for “plan mode” or special tools – just say it in words.

### Give feedback tools

Claude Code iterates itself if it has a way to verify the result. Give him:

• unit tests
• integration tests
• screenshots via Puppeteer
• screenshots of the iOS simulator

For example, give a layout and say “build this web interface.” After 2-3 iterations with visual feedback, the result is almost perfect without you.

### Use voice input

On macOS: System settings → Special features → Dictation. Double press the dictation button and you can talk. Proms can not be printed, but spoken out loud, as a task to a colleague.

### Useful incantations

commit, push # Claude will look at the git-log, select the commit format and start
explore [task], make a plan # exploration + plan without writing code
use [tool] -help to learn how # teach Claude to work with your CLI tool

## Memory and context: CLAUDE file. md

For Claude to remember the important things between sessions, create a CLAUDE.md file at the root of the project. It automatically loads into context at each launch.

What's in there

• Frequent bash teams of the project
• MCP tools used
• Architectural solutions
• Important Files You Need to Know

Keep the file short: too long CLAUDE.md just eats context.

** Memory file hierarchy:**

• CLAUDE.md at the root of the project - common to the entire team, commits in git
• CLAUDE.md in nested directories - pulls up automatically when Claude is working in this directory
• Local CLAUDE.md - only for you, does not commit
• Enterprise level – global config for all employees

Check which memory files are currently downloaded: /memory

## Team Tools: Bash and MCP

When the basic scenarios are mastered, connect the tools of your team.

Bash Tools: Just tell Claude you have a CLI. For example:

We have a barby-cli tool. Use barby-cli-help to figure out how to use it, and apply it to it.

MCP tools: Claude Code supports MCP out of the box. Add an MCP server, explain how to use it, and Claude will start using it. An example from Anthropic: In general, the Puppeteer MCP server has been tested and any team engineer automatically accesses it without manual installation.

Enterprise-policies: through the config you can centrally:

• Automatically allow specific bash teams for all employees
• Block certain URLs or commands
• Connect MCP servers for the entire team at once

## Hot keys that few people know

## Claude Code SDK and Use in Pipelines

Claude Code is built on the same SDK that is available to everyone. This means that it can be built into any piplins.

# Simple example: CLI, output to JSON
claude -p "your promt" --allowedTools bash:mycommand --output-format json

# Piping: transfer git status and process the result
git status | claude -p "What's important here?" | jq '.result'

As used in Anthropic:

• In CI/CD Pipelines
• For incident response
• To read logs from GCP buckets and analyze anomalies
• To download data from Sentry CLI and analyze them

According to Boris, Think of it as a smart Unix utility. You give me a prompt, you get JSON. The combinations are endless.”

## Parallel work (advanced level)

Power users run multiple copies of Claude simultaneously in different branches of the same repository. git worktrees is used for isolation. This allows you to solve several tasks in parallel.

The Anthropic team is working to simplify this workflow, but it’s already a working approach for maximum performance.

## Multimodality

Claude Code is fully multimodal, although this is not obvious in the terminal. You can:

• Drag the image directly to the terminal
• Indicate the path to the image file
• Copy and paste the image

Typical scenario: give a screenshot of the layout → ask to collect UI → Claude iterates from Puppeteer screenshots.

## Recommended learning path

1. Start with Q&A – ask questions about your codebase, don’t ask to write code right away
2. Move to tasks with a plan – add “plan first” to any implementation request
3. Give feedback tools – tests, screenshots, anything for iteration
4. Configure CLAUDE.md - once, for the whole team
5. Connect command tools – CLI, MCP servers, bash commands
6. Explore SDK - embed in Pipelines and CI

According to Anthropic, about 80% of the company’s technicians and researchers use Claude Code every day. *