Claude Code power tips: loops, hooks, and remote workflows

A recent thread from Boris Cherny spotlights under-used Claude Code features that make it feel more like a programmable teammate. Highlights include session teleporting, remote control, hooks, and automation via /loop and /schedule—plus a reminder that usage limits can get in the way.

claude cover

TL;DR

  • Mobile + session mobility: iOS app Code tab; claude --teleport / /teleport to continue sessions across devices
  • Remote control: /remote-control to drive a local session from phone/web; enable for all sessions via /config
  • Automations: /loop and /schedule run periodic workflows up to a week; examples include babysit, Slack feedback, post-merge sweeps
  • Hooks: SessionStart, PreToolUse, PermissionRequest, Stop to add deterministic lifecycle logic and organizational memory
  • Verification + parallelism: Chrome/Edge extension and Desktop app browser testing; /branch, /btw, git worktrees (claude -w), /batch for large migrations
  • Scripting knobs + constraints: --bare, --add-dir, --agent, /voice, Cowork Dispatch; replies cite early usage limits and MacBook fan noise

A recent thread from Boris Cherny collects a surprisingly practical set of “hidden” and under-used features in Claude Code—less about clever prompts, more about treating the tool like a programmable collaborator that can move between devices and keep working on its own.

The list spans mobile, desktop, and terminal workflows, but the throughline is consistent: give the agent ways to verify work, keep state across contexts, and automate the repetitive glue around shipping code.

Mobile, session mobility, and remote control

Cherny starts with a reminder that Claude Code work isn’t confined to a laptop. Claude’s mobile app includes a Code tab, and Cherny notes using the iOS app to make changes without opening a computer. The thread links to the app download page: https://t.co/WiuGrbfeyI.

From there, the emphasis shifts to continuity and control:

  • Teleporting sessions between mobile/web/desktop and terminal can be done with claude --teleport or /teleport to continue a cloud session locally. (https://t.co/mJlFQdjm50)
  • Remote-controlling a local session from a phone or the web is available via /remote-control. Cherny mentions enabling remote control for all sessions via /config.

In the replies, one developer calls /remote-control “criminally underused” for async workflows—directing a local session running on a Mac mini from a phone.

Automations: /loop, /schedule, and “skills” that run themselves

Two features Cherny highlights as especially powerful are /loop and /schedule, which can run Claude automatically at set intervals for up to a week.

Cherny shares examples of local loops used to keep PRs moving:

  • /loop 5m /babysit to address code review, auto-rebase, and shepherd PRs
  • /loop 30m /slack-feedback to post PRs for Slack feedback
  • /loop /post-merge-sweeper to raise PRs for missed review comments
  • /loop 1h /pr-pruner to close stale PRs

The suggestion is to experiment with turning workflows into “skills” and then attaching them to loops. Source link: https://t.co/cx6YMibGN3.

A reply echoes this approach, describing teams running a babysit loop on client projects to catch review comments, rebases, and small regressions.

Hooks: deterministic logic in the agent lifecycle

Hooks are presented as the way to inject deterministic behavior into Claude Code’s lifecycle—less “chat,” more repeatable plumbing.

Examples Cherny lists include:

  • SessionStart for dynamically loading context when Claude starts
  • PreToolUse for logging every bash command the model runs
  • PermissionRequest for routing permission prompts to WhatsApp for approval/denial
  • Stop for prompting Claude to continue whenever it stops

Documentation is linked here: https://t.co/GukTJ9iy0R. In the replies, one developer notes a SessionStart hook that loads product context—customer problems, decision history, architectural constraints—so sessions begin with “organizational memory.”

Verification loops for frontend: extension + Desktop app testing

For frontend work, Cherny argues the key is giving Claude a way to verify output so it can iterate until it looks right.

Two options are called out:

  • The Chrome/Edge extension, which Cherny says is used “every time” for web work and tends to be more reliable than other similar MCPs. (https://t.co/m7wwQUmp1C)
  • The Claude Desktop app, which bundles the ability for Claude to automatically start a web server and test it in a built-in browser. (https://t.co/KYD9S6GY4X)

Branching, side-queries, and parallelism at scale

A few features focus on keeping work moving without breaking flow:

  • Forking a session via /branch, or from the CLI with claude --resume <session-id> --fork-session. (https://t.co/zqj1Iwh361)
  • /btw for side queries, used for quick questions while an agent continues working. (https://t.co/hO02jUTWh2)

Then there’s the more industrial-strength parallelism:

  • git worktrees support, with claude -w to start a new session in a worktree (or a worktree checkbox in the Desktop app). Cherny notes using worktrees to run “dozens of Claudes” in the same repo, and mentions a WorktreeCreate hook for non-git VCS users. (https://t.co/Yv3SlQr1Kr)
  • /batch, which “interviews” the user and then fans work out to as many worktree agents as needed—“dozens, hundreds, even thousands”—aimed at large migrations and other parallelizable changes. (https://t.co/grOtCuG8Wz)

SDK and configuration knobs: --bare, --add-dir, and custom agents

The thread also digs into flags that matter once Claude Code becomes part of scripts and multi-repo work:

  • --bare: Cherny says default startup for claude -p (and the TypeScript/Python SDKs) searches for local CLAUDE.md files, settings, and MCPs; for non-interactive usage it’s often better to specify inputs explicitly via --system-prompt, --mcp-config, and --settings. Cherny calls the default behavior a design oversight and says a future version will flip the default to --bare—but for now it’s opt-in.
  • --add-dir (or /add-dir): used to give Claude access (and permissions) to more folders, useful across multiple repositories. There’s also a mention of additionalDirectories in a team settings.json. (https://t.co/HP12BtgFLv)
  • --agent: custom agents can be defined in .claude/agents and run with claude --agent=<name>. Cherny later confirms in replies the point can be non-interactive mode or customizing the main loop. Links: https://t.co/BqVPwRthLv and https://t.co/JiB98KI5eA

Dispatch and voice input

Two final workflow pushers:

  • Cowork Dispatch is described as a secure remote control for the Claude Desktop app, able to use MCPs, browser, and computer with permission—used for catching up on Slack/email, file management, and laptop tasks when away from the computer. (https://t.co/6nuhsh8Gg0)
  • /voice enables voice input in the CLI; Cherny says most coding is done by speaking rather than typing. /voice plus holding space in CLI, using the voice button on Desktop, or enabling dictation on iOS are mentioned.

The friction point in the replies: usage limits

While the feature list is expansive, several replies focus on practical constraints: multiple commenters say Claude Code is “getting exhausted too soon,” hitting usage limits early in the day even on paid or “Max” plans, and one mentions increased MacBook fan noise.

Those comments sit in contrast to the thread’s core idea: Claude Code has a deep bench of automation and parallelism tools—but they’re most valuable when sustained usage is available.

Original source: https://x.com/bcherny/status/2038454336355999749

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community