Claude Code Goes Full‑Stack: How the New ‘Computer Use’ Feature Automates Development

Anthropic has released a major update to Claude Code by adding a "computer use" capability, allowing the model to act like a real programmer: it can write code, modify files, run commands, launch applications, and even perform UI interactions such as clicking and scrolling.

Auto‑Driving Mode

With a single /mcp command, users can enable an "auto‑driving" mode where Claude takes full control of the Mac, operating the browser, mouse, keyboard, and screen to complete end‑to‑end tasks:

Cross‑application interaction : open installed apps and perform UI actions.

End‑to‑end loop : write code → compile → launch app → UI automation → detect bugs → fix code → verify.

Tool‑agnostic execution : works with SwiftUI, Electron, or any GUI tool without a CLI.

This creates a complete development loop that requires no manual intervention, dramatically reducing the cost of switching between tools.

Token‑Cost Bugs and Temporary Workarounds

Shortly after launch, users discovered two severe bugs that cause cache invalidation and inflate token consumption by 10‑20×, effectively exhausting quota. The first bug originates from a string‑replacement issue in the Bun runtime; a temporary fix is to run Claude via npx @anthropic-ai/claude-code. The second bug crashes the cache when using the --resume flag; the only current mitigation is to revert to an older version with reduced functionality.

Boris Cherny’s 15 Hidden Tips

Claude Code’s creator, Boris Cherny, shared a list of advanced commands that unlock hidden capabilities:

Embedded compiler : use Claude on iOS to write and submit code directly from a mobile device.

Cross‑device teleportation : /teleport moves a cloud session to a local terminal; /remote-control lets a phone or web client control a local session.

Automation loops : schedule recurring tasks with /loop 5m /babysit, /loop 30m /slack-feedback, /loop /post-merge-sweeper, /loop 1h /pr-pruner.

Lifecycle hooks : SessionStart loads context automatically; PermissionRequest pushes approvals to WhatsApp.

Remote commander ( Cowork Dispatch) handles non‑coding chores like Slack messages or file management.

Frontend eye : attach a Chrome extension so Claude can see and iterate on front‑end output.

Server auto‑start and test : Claude Desktop can launch a web server and run tests in its built‑in browser.

Forked sessions : /branch creates a parallel conversation; CLI equivalent is claude --resume <session-id> --fork-session.

Git worktrees : claude -w enables parallel work across multiple worktrees within the same repository.

Batch processing : /batch interviews the user for intent, then spawns hundreds of worktree agents to execute large‑scale migrations.

Bare mode : --bare skips configuration scans, speeding up SDK startup tenfold.

Directory expansion : --add-dir grants Claude access to additional folders for cross‑repo collaboration.

Custom agents : define read‑only or audit agents in .claude/agents/ and invoke them with claude --agent=xxx.

Voice input : /voice lets users dictate code by holding the space bar or tapping a voice button.

Continuous operation : combine these features to turn Claude from a helper into a 24/7 engineering team.

References

For further details see the original X posts, Reddit discussion of the cache bugs, and the GitHub best‑practice guide linked in the article.