Source: raw/Mastering_Claude_Code_in_30_minutes.md (YouTube 6eBSHbLKuN0, official Anthropic channel, published 2025-05-22, 1.69M views at fetch; transcript from the auto-captions of an X re-upload, x.com/dravenip/status/2114399052259287226, posted 2026-10-25)

Creator: Boris Cherny (creator of Claude Code), Anthropic · URL: https://www.youtube.com/watch?v=6eBSHbLKuN0 · Duration: 28:07 · Platform: YouTube (Code with Claude conference session, 2025-05-22)

Boris Cherny’s practical-tips session from Anthropic’s first Code with Claude conference. He gave it on the day Anthropic announced the Claude GitHub app, and an SDK session ran across the hall at the same time. It is a beginner-to-power-user ladder: start with codebase Q&A, then editing with a plan, then your team’s tools, then context through CLAUDE.md and shared settings, then headless claude -p and parallel sessions. The ladder still holds. Several specific mechanics have changed since, and those are flagged below. The X re-upload that surfaced it in October 2026 called it a newly shared workshop on “how to actually prompt Claude”. It is neither new nor a prompting workshop. It is a 17-month-old Claude Code talk.

Key Takeaways

  • Start with codebase Q&A, not code edits. Anthropic teaches this on new hires’ first day. Boris says technical onboarding went from “about two or three weeks” to “about two or three days.” Q&A also teaches people what Claude Code can one-shot and where it needs guidance.
  • No indexing, no upload. Claude Code explores the code with tools at question time, so there is no remote index, no setup wait, and the code stays local.
  • Git history is a question you can ask. For example, “why does this function have fifteen arguments?” makes it dig through the log, the commits and their linked issues. Boris uses “what did I ship this week?” for his Monday standup.
  • Plan before you build, in plain words. “Before you write code, make a plan” works without plan mode. Asking for brainstorming and approval first is his fix for big features that come back wrong.
  • Give it a way to check its own work. Unit tests, Puppeteer screenshots or the iOS simulator let it iterate two or three times. In his account that takes a UI built from a mock from “pretty good” to “almost perfect”.
  • Shared project context gives the most leverage. A checked-in CLAUDE.md, a checked-in .mcp.json and shared permissions are written once and benefit the whole team. Keep CLAUDE.md short: common bash commands, the style guide, core files.
  • Treat claude -p as a Unix utility. Prompt in, JSON out, pipeable both ways. Anthropic used it in CI, incident response and log triage.
  • Power users run sessions in parallel, over SSH and tmux, with several checkouts or git worktrees. Boris called himself a “Claude normie” who usually runs one session.

The Ladder

1. Set up the environment

  • /terminal-setup for Shift+Enter newlines, /theme (light, dark, or colorblind-friendly “Daltonize” themes), /install-github-app to @-mention Claude on issues and PRs.
  • Allowlist the tools you approve often so you aren’t prompted for them every time.
  • Dictate prompts. Boris turns on macOS dictation (System Settings > Accessibility) and speaks his prompts. Specific prompts work better, and speaking makes long specific ones cheap.

2. Codebase Q&A

  • Example prompts: how is this code used, how do I instantiate this class, why was this written this way (git history), and what does this GitHub issue say (fetched with WebFetch).
  • It goes “a level deeper” than text search, finding real instantiation and usage examples.
  • Git skill is not system-prompted: “It knows it because the model was awesome.”

3. Editing code

  • The toolset is small (edit files, run bash, search files) and the model chains the tools without being told which to use.
  • “Commit, push, PR” is a phrase he uses: it creates a branch, matches the repo’s commit format from git log, pushes and opens the PR.
  • Three workflows: explore → plan → confirm → code; write tests and iterate until they pass; give it a visual mock plus a screenshot tool and let it iterate.

4. Your team’s tools

  • Bash tools: tell it about the CLI and that it can read --help. Put frequently used tools in CLAUDE.md so every session knows them.
  • MCP tools: add the server and tell Claude how to use it.

5. Context

MechanismWhereNotes from the talk
CLAUDE.mdProject rootRead into the first user turn of every session. Check it in and share it
CLAUDE.local.mdProject rootPersonal and not checked in
Nested CLAUDE.mdChild directoriesPulled in when Claude works in that directory
Enterprise CLAUDE.mdManaged rootShared across all of a company’s codebases
Slash commands.claude/commands/, project or homeAn issue-labelling slash command runs in a GitHub Action and labels new issues in the Claude Code repo
@-mentionsInlinePull specific files into context
  • Settings follow the same hierarchy (project, project-local, global, enterprise policy) for slash commands, permissions and MCP servers. Enterprise policy can auto-approve a test command for every employee or block a URL so it can never be fetched, and employees can’t override it.
  • /memory shows and edits every memory file in play.
  • He suggests running your context through a prompt improver, and deciding for each item who it is for and whether it loads every time or on demand.

6. Keyboard shortcuts (as of May 2025)

KeyAction
Shift+TabAuto-accept edits (bash still asks)
#Save something to memory (CLAUDE.md)
!Bash mode. The command and its output go into context
@Mention files and folders
EscStop at any point. “It’s not gonna corrupt the session”
Esc EscJump back in history
--resume / --continueResume a past session
Ctrl+RShow full output, the same text Claude sees in its context

7. Headless and parallel

  • claude -p "<prompt>" with a list of allowed tools (specific bash commands included) and an output format of JSON or streaming JSON. Current flag names are in the CLI reference. Pipe git status into it, pipe the output through jq, read a large GCP-bucket log or Sentry CLI output and ask what’s interesting.
  • At Anthropic this ran in CI, incident response and pipelines. “It’s like a super intelligent Unix utility… we barely scratched the surface.”
  • Parallel work: multiple checkouts or git worktrees, sessions over SSH/tmux. He said Anthropic was “actively working on making this easier.”

Q&A Highlights

  • Hardest part to build: making bash safe without approval fatigue. The solution was read-only command detection, static analysis of which commands combine safely, and a tiered allow/block permission system, because not everyone runs code in a container.
  • Multimodal from day one: drag and drop an image, paste it, or give a file path. Boris drops in a design mock and adds a Puppeteer MCP server so Claude can iterate against it.
  • Why a CLI, not an IDE: Anthropic engineers use every editor (VS Code, Zed, Xcode, Vim, Emacs), so the terminal is the common denominator. He also predicted “a good chance that by the end of the year people aren’t using IDEs anymore”, so they avoided over-investing in UI.
  • Adoption (May 2025): “about 80% of people at Anthropic that are technical use Claude Code every day”, researchers included, using the notebook tool.

What Has Changed Since May 2025

The wiki documents these changes to mechanics shown in the talk:

  • The # memory hotkey is retired. Anthropic now says it “used to” encourage #, and Claude saves relevant memories automatically. See the CLAUDE.md primer (superseded note dated 2026-07-24) and Auto Memory.
  • Esc Esc now opens the /rewind picker (/undo is an alias since W16). See What’s New W16 and Claude 5.5 mistakes and fixes.
  • The SDK he demos is now the Claude Agent SDK. claude -p remains the headless entry point. See Claude Agent SDK for the rename and the 2026 billing history of programmatic usage.
  • Install is no longer Node-only. He said “all you need is Node.js.” Native binaries replaced bundled JavaScript in v2.1.113 (W16).
  • Plan-first is no longer the default advice. By June 2026 Boris and Cat Wu said auto mode had replaced plan mode for current models (Reflecting on a Year of Claude Code). In February 2026 he still recommended plan mode for about 80% of tasks (Lenny’s Podcast).
  • AGENTS.md is read when no CLAUDE.md exists (v2.1.277), an addition to the context hierarchy shown here. See the CLAUDE.md primer.

Try It

  1. Onboard someone with Q&A only. Day one, no edits: have them ask about architecture, the oldest strange function (through git history), and “what did I ship this week?”. This is Anthropic’s own onboarding method from the talk.
  2. Check in the shared layer. Commit a short CLAUDE.md (bash commands, style rules, core files) and an .mcp.json with the servers the team uses. Boris names this as the starting point when you’re unsure which configuration to use.
  3. Add a feedback loop to one UI task. Give Claude a mock plus a screenshot tool (Puppeteer or Chrome MCP) and require two or three self-check passes before it reports done.
  4. Script one recurring triage job with claude -p. For example, <log source> | claude -p "what's unusual here?" --output-format json | jq ... in a cron job or CI step. Check the current programmatic-usage billing in Agent SDK first.
  5. Skip the # hotkey from the talk. Say “save this to memory” or rely on auto-memory.

Open Questions

  • The transcript comes from machine captions of a re-upload. The on-screen slides (the exact install command and the settings-hierarchy matrix) are not in the transcript, and the details in the Context section come from his spoken description of them.
  • The “two to three weeks to two to three days” onboarding figure and the 80% daily-use figure are Boris’s statements on stage, not published metrics.