Most Claude Code users type prompts into a terminal and wait. Boris Cherny, the head of Claude Code at Anthropic, apparently writes most of his code by talking to it from his phone. That gap between typical usage and what's actually possible is the subject of his latest tips thread, posted March 30 - the fifth in a series that now totals over 40 recommendations.
Here are the features worth paying attention to, grouped by how they change your workflow.
Code From Anywhere
Claude Code has a mobile app. Open the Claude app on iOS or Android, tap the Code tab, and you can review changes, approve PRs, and write code from your phone. Cherny says he uses this regularly.
The /teleport command moves a cloud session to your local terminal. /remote-control lets you steer a locally running session from your phone or browser. If you've ever started something on your laptop and wanted to check on it from the couch, this is the answer.
Voice input works too: /voice in the CLI, then hold spacebar. Cherny says he does most of his coding this way. On the Desktop app there's a dedicated voice button.
Parallel Everything
claude -w starts a session in a git worktree (an isolated copy of your repo). Cherny says he runs "dozens of Claudes at all times" this way, each working on separate tasks without stepping on each other.
/batch takes this further. It interviews you about what needs doing, then fans out the work to as many worktree agents as needed - dozens, hundreds, even thousands. Large code migrations that would take days of sequential work become parallel.
/branch forks your current conversation, so you can explore a different approach without losing your place. To resume the original, use claude -r with the session ID.
Automation Loops
/loop and /schedule let you run commands on a recurring interval, up to a week at a time. Cherny's examples reveal how he actually uses these:
/loop 5m /babysit- auto-address code review feedback, auto-rebase, shepherd PRs to production/loop 30m /slack-feedback- automatically create PRs based on Slack feedback/loop 1h /pr-pruner- close stale PRs
The pattern: turn a workflow into a custom skill, then put it on a loop.
Developer Efficiency
Hooks run deterministic logic at specific points in Claude's lifecycle. SessionStart can load dynamic context. PreToolUse can log every bash command. PermissionRequest can route approval prompts to WhatsApp. Stop can poke Claude to keep going whenever it pauses.
--bare flag speeds up SDK startup by up to 10x. By default, claude -p searches for CLAUDE.md files, settings, and MCP servers. For non-interactive scripts, --bare skips all of that. Cherny calls the current default "a design oversight" and says they plan to flip it in a future version.
--add-dir gives Claude access to additional folders when working across multiple repos. You can also set additionalDirectories in settings.json to always load specific paths.
--agent loads custom agent definitions from .claude/agents/ with restricted tools, custom system prompts, and specific models. Useful for read-only review agents or domain-specific assistants.
/btw lets you ask a side question without interrupting the current task. Claude answers in a dismissible overlay and keeps working.
One tip Cherny emphasized above the rest: the Chrome extension for frontend work. His argument is simple - asking someone to build a website without letting them use a browser won't produce good results. The extension lets Claude verify its own output and iterate until it looks right. He says it works more reliably than alternative MCP-based approaches.
These aren't experimental features buried in a dev branch. They're shipping in the current version of Claude Code. The gap between what most people use and what's available is wide enough that revisiting the feature set is worth the time.