Claude Code Creator Boris Talk Summary: 14 Practical Tips You Might Not Know

Who This Talk Is By and What It Covers

The speaker is Boris, Anthropic Member of Technical Staff and creator of Claude Code. In a 28-minute talk, he completely skips theory and historical background and directly demonstrates “how to use Claude Code day-to-day.”

The value of this talk is that Boris covers a bunch of details—the # key auto-writes to CLAUDE.md, Shift+Tab toggles auto-accept, nested CLAUDE.md directories auto-load, claude -p as a Unix utility to pipe into pipelines—details that might never show up in docs, needing the creator to demo personally.

I’ve organized the talk highlights into 14 points, grouped by logic, plus 6 takeaways for developers at the end.

1. What Claude Code Is

It’s not a “completion tool”—it’s a fully agentic AI assistant.

Not the kind that fills in code line-by-line for you—it builds features, writes entire functions, fixes entire bugs.

The best part: it doesn’t conflict with your existing IDE workflow. VS Code, Xcode, JetBrains, Vim, Emacs all work simultaneously—because Claude Code runs in terminal, SSH, tmux, anywhere.

2. Initial Setup (Do These 4 Things After Install)

1. /terminal-setup → Set Shift+Enter for newline (no backslash needed)
2. /theme → Set light / dark / colorblind-friendly mode
3. /install-github-app → Mention @claude directly in GitHub issues/PRs and it’ll respond
4. macOS: Enable Dictation in System Settings, double-click to enable voice input—speaking prompts is much faster than typing

Boris mentioned most of his prompts are spoken—they’re faster, especially for long prompts or describing requirements.

3. Don’t Rush to Code—Use Codebase Q&A First

Boris’s strongest recommendation for beginners: ask codebase questions first, don’t rush to have Claude write code.

Example questions:

• “How do I use this function?”
• “Why does this function have 15 parameters? What’s the historical context?”
• “Where is this class instantiated?”

Claude doesn’t just do text search—it reads git history, traces GitHub issues, and connects the dots to give you the full context.

Anthropic’s internal onboarding reduced from 2-3 weeks to 2-3 days—because new hires don’t need to bother senior engineers, they just ask Claude about the codebase.

Also: code isn’t uploaded, no indexing happens, and they won’t use your code to train models. Just open and use—no waiting for indexing.

4. Weekly Standup Hack

Boris shared that before every Monday standup, he asks Claude:

“What did I ship this week?”

Claude automatically reads git log, knows your username, gives you a nice readout—just copy-paste into the doc and you’re done.

Works for any team using git—weekly reports, monthly reports, quarterly reports all apply.

5. Plan Before Coding (No Need for Plan Mode)

The simplest and most effective habit:

“First make a plan, show me, then write code.”

No need to switch to plan mode, no special tool needed, just say it like that. Claude will brainstorm, create a plan, wait for your confirmation before moving.

Saves you from the disaster of “wrote 3000 lines in the wrong direction.”

Another common incantation:

“commit push PR”

Claude itself reads your git log to learn your commit format, opens a branch, pushes to GitHub, opens a PR. No explanation needed.

6. Connecting Team Tools—bash CLI + MCP

There are two ways to connect your team’s tools to Claude Code:

bash CLI: Just tell Claude “use this CLI, run --help and you’ll figure it out.”

MCP servers: Same idea—tell it what MCPs are available and it’ll use them.

Pro tip: Just write commonly used ones into CLAUDE.md (covered later)—it’ll remember across sessions.

7. Give Claude Self-Verification Ability → It Will Iterate and Improve

This is one of the most valuable concepts from the entire talk.

Just give Claude a tool to check its own work—it will iterate 2-3 times and become nearly perfect.

• Writing frontend? Give it puppeteer screenshots
• Writing iOS App? Give it simulator screenshots
• Writing backend? Give it unit tests / integration tests
• Any domain? Just give it a feedback loop

Boris emphasized that one-shot results are usually not perfect, but with a feedback loop, iterating 2-3 times jumps quality up by an order of magnitude.

8. CLAUDE.md: The Most Important Context Tool

Placed in project root, CLAUDE.md gets automatically loaded into every session.

What to put in it:

• Your commonly used bash commands
• Common MCP tools
• Important architecture decisions
• Core file locations
• Standards and style guides

What to note:

• ⚠️ Keep it brief. Writing too long wastes context window and actually hurts.
• You can also put CLAUDE.md in nested directories—Claude auto-loads when working in that directory.
• Big companies can put it at enterprise root—shared across the whole company.

Common config is ~/.claude/CLAUDE.md (personal global) + per-project CLAUDE.md (project level)—most people haven’t explored the nested directory part—it gets way more precise for large monorepos subdivided into subdirectories.

9. Slash Commands (Your Own Common Routines)

.claude/commands/ goes in home or project root—you define your own slash commands.

Example from Boris: The Claude Code repo itself has a slash command called “label GitHub issues,” paired with a GitHub Action that runs automatically—issue auto-labels, no human work needed.

Super useful for teams: Prompts you find yourself repeating, write them as slash commands—just call one word next time.

10. Hierarchy Levels (Careful Configuration)

Claude Code settings have three levels:

1. Project level (checked into git repo, shared across the whole team)
2. User level (personal ~/.claude/)
3. Enterprise level (company-enforced policy)

Scope: commands / permissions / MCP servers / Bash whitelist / blacklist.

Enterprise level applications:

• Force block specific URLs (employees can’t override)
• Auto-approve specific common commands company-wide (no need to ask every time)
• Unified MCP servers configuration

11. Terminal Hidden Key Bindings (Biggest Talk Surprise)

These keybindings many people used for half a year without discovering:

The # key is especially useful: if you often tell Claude “don’t do it that way next time,” just write it into CLAUDE.md with #—it’ll remember next time without you having to repeat.

12. Claude Code SDK = claude -p

Many people don’t know: the claude -p flag you’ve been using is the SDK.

Usage:

• CI/CD pipeline automation
• Incident response auto-analysis
• Any Unix pipeline chaining

Example:

1

git status | claude -p "analyze these changes for me, give me a PR description" --format json

Boris called it: “super intelligent Unix utility.”

Chaining claude -p into existing shell pipelines and using jq to process JSON output is a pattern not widely adopted yet but incredibly powerful—any CI/CD, cron job, incident response flow can use it.

13. Power User Mode (Multiple Claudes in Parallel)

Boris observed: Anthropic’s power users don’t use one Claude—they run multiple simultaneously.

Tricks:

• Multiple SSH sessions + TMUX panes
• Multiple checkouts of the same repo
• git worktree isolation + multiple Claudes in parallel
• Running “writing Claude” + “reviewing Claude” + “testing Claude” simultaneously

This pattern is especially useful for “one person wearing multiple hats” workflows—running development, review, and testing three Claudes in parallel is much faster than running them one by one.

14. Q&A Highlights

What’s the hardest part?

Bash command security. Manually approving each one is tedious, auto-approving everything is dangerous. Solutions:

1. Distinguish read-only commands
2. Static analysis to see if command combinations are safe
3. Tiered permission (allow list, block list)

Multimodal support?

Drag/drop images, give file paths, copy-paste images—all three supported.

Boris’s common demo: give Claude a mock image + puppeteer—it compares the screenshot vs mock and iterates until they match.

Why CLI instead of IDE?

Boris’s answer has two layers:

1. Hard to get consensus inside Anthropic on which IDEs to use
2. “Personally I think I might stop using IDEs by year-end”—he doesn’t want to over-invest in UI layer

Point 2 is worth thinking about—if models keep getting stronger, IDEs might indeed not be the primary interface anymore.

Claude Code in ML field?

~80% of Anthropic’s technical staff use Claude Code daily, including researchers using the notebook tool to directly edit notebooks.

6 Takeaways for Developers

1. Shift+Tab toggles auto-accept — Once you trust Claude, stop manually approving every time—it saves tons of context switching
2. # memory hack — When telling Claude “don’t do it that way next time,” just use # to write to CLAUDE.md—no more repeating
3. Slash commands — All prompts you use repeatedly, write them as slash commands—just call one word next time
4. claude -p pipe chaining — Treat Claude as a Unix utility—chain it into CI/CD, cron, incident response flows
5. Nested CLAUDE.md — Global + project level + subdirectory level—three-tier config; large monorepos benefit especially
6. Multiple Claudes in parallel — Use git worktree isolation—run writing Claude / reviewing Claude / testing Claude simultaneously

Closing

The value of this talk isn’t in qualitative descriptions like “how powerful Claude Code is”—it’s in the details Boris personally demonstrated—# key, Shift+Tab, nested CLAUDE.md, claude -p pipe. These details might never show up in docs—only when the creator personally says “this is how I use it day-to-day” can it be passed to the community.

If you’re using Claude Code, the full 28 minutes are worth watching from start to finish.

Link: Boris on X (original thread)