Claude Code updates hidden-features guide with /loop, hooks, and /batch

• Anthropic's Boris Cherny turned a grab-bag of "hidden and under-utilized features" into a concrete Claude Code workflow guide, covering mobile coding, session handoff, scheduled jobs, hooks, browser verification, worktrees, /batch, and CLI flags that change runtime behavior feature guide.
• The biggest implementation shift is that Claude Code is being framed as a cross-device control plane: Cherny says he writes code from the iOS app mobile app, then moves sessions between cloud and local machines with /teleport and /remote-control session handoff.
• The guide also treats Claude Code as a persistent automation surface. Cherny's examples use /loop and /schedule for recurring PR and Slack workflows scheduled loops, while hooks can run deterministic logic at lifecycle events like SessionStart, PreToolUse, and PermissionRequest hooks.
• For larger repos and SDK usage, the thread highlights parallelism and startup tuning: worktree-native sessions worktrees, /batch fan-out for migrations batch, multi-repo access via --add-dir add-dir, custom agents via --agent custom agents, and --bare to cut SDK startup by "up to 10x" in non-interactive runs bare mode.

Cherny's thread is not a product launch so much as a map of features that already exist but are easy to miss. The practical change is packaging: instead of scattered docs, one maintainer-level post now ties together mobile, remote control, scheduled tasks, hooks, browser tooling, worktrees, and agent configuration into a single operating model for Claude Code feature guide.

That matters because the tool's surface area has grown fast enough that users are building their own references. A recent Hacker News thread around a community cheat sheet describes Claude Code as changing often enough that a quick reference saves users from "digging through docs" HN thread, and the cheat sheet itself tracks slash commands, MCP setup, memory files, CLI flags, and recent changes such as conditional hooks cheat sheet.

The mobile angle is one concrete example. Cherny says "I write a lot of my code from the iOS app" through the app's Code tab mobile app. That is paired with cross-device continuity rather than a separate mobile-only mode.

The cross-device model has two paths. Cherny says claude --teleport or /teleport can continue a cloud session on your machine, while /remote-control lets a phone or browser control a locally running session session handoff. Anthropic's remote-control docs add the implementation detail that the session keeps running on the user's machine, with access to local files, tools, and project config rather than moving that context into the cloud.

Automation is the other half of the guide. Cherny calls /loop and /schedule "two of the most powerful features" and says he runs loops for /babysit, /slack-feedback, /post-merge-sweeper, and /pr-pruner to handle rebases, code review follow-up, and stale PR cleanup scheduled loops. The docs behind scheduled tasks separate session-scoped local loops from cloud tasks, which is a meaningful deployment distinction if the job needs local file access.

Hooks push this further into programmable agent behavior. Cherny's examples include loading context on SessionStart, logging every bash command on PreToolUse, routing permission prompts to WhatsApp on PermissionRequest, and nudging the model to continue on Stop hooks. For browser-facing work, he says the most important tip is to give Claude "a way to verify its output," and recommends the Chrome extension for frontend tasks because it "tends to work more reliably" than similar MCP approaches Chrome extension. Dispatch extends the same remote-control idea to the desktop app, where Cherny says it can manage Slack, email, files, and MCP-backed tools from away from the laptop Dispatch.

The guide's scaling primitives are worktrees, fan-out, and narrower process startup. Cherny says Claude Code has "deep support for git worktrees," using claude -w or the desktop worktree checkbox to run many parallel sessions in one repository; he adds, "I have dozens of Claudes running at all times" worktrees. /batch sits on top of that model: he says it interviews the user, then fans work out to as many worktree agents as needed, including "dozens, hundreds, even thousands," for migrations and other parallelizable changes batch.

Repository scope and agent scope are also now first-class knobs. --add-dir or /add-dir grants access to additional folders for multi-repo work, and the same setting can be made persistent with additionalDirectories in team settings add-dir. --agent lets teams define custom agents in .claude/agents with their own system prompts and tool access custom agents, which Anthropic's subagent docs position as a way to preserve context and route tasks to specialized assistants.

For SDK users, Cherny also flags a performance gotcha: default startup still searches for local CLAUDE.md, settings, and MCPs, but --bare skips that implicit loading for non-interactive runs. He says this can speed SDK startup by "up to 10x," and says the current default was a "design oversight" that Anthropic plans to flip in a future release bare mode.