/
Published on 14.03.2026
12 Claude Code updates shipped this week — Multitasking with /btw, recurring task scheduling with /loop and desktop tasks, improved auto-memory, voice prompting, Excel/PowerPoint sync, revamped Skills Creator with evals, multi-agent code review, and effort levels. Major structural changes, not incremental patches.
The updates aren't cosmetic. They change how Claude reasons, what it remembers, how it schedules work, and how it reviews code. Here's what shipped:
/btw Command (Interruption-Free Queries)You're mid-refactor. Claude is working through a complex change. You need a quick syntax detail but don't want to derail the current job.
/btw what is the syntax for a React hook? Claude answers in a dismissible overlay. The response never enters the conversation history. Your context window stays clean. The main task keeps running in the background.
Why it matters: Multitasking within a single Claude Code session. Quick answers without context pollution. This is how human collaborators work—quick sidebar conversations without derailing the main work.
/loop Command (Recurring Task Scheduling)/loop 15m check build runs every 15 minutes.
/loop audit deps every 2 hours parses natural language into a cron schedule.
/loop 1h /security-review re-runs a packaged skill hourly.
Three guardrails: tasks expire after 72 hours, no catch-up runs if your machine sleeps, everything is session-scoped (close the terminal, the schedule disappears).
Why it matters: Status monitoring, dependency checks, log scanning—all running while you focus on code. The practical example: automatically generate summaries of recently merged PRs at set intervals.
Unlike /loop (temporary, terminal-scoped), desktop tasks in the Claude Desktop app persist through reboots. Create a task once; it runs daily/weekly/hourly as long as your computer is awake and the app is open.
Use cases: Daily code reviews, dependency update checks, morning briefings pulled from your calendar or inbox.
Wire desktop scheduled tasks to send output to a Telegram bot. Every time the task runs, the result lands in your Telegram chat: team alerts, build notifications, daily summaries.
Setup: Ask Claude to add a Telegram messaging module, create a bot via BotFather, add credentials to .env, end your prompt with "send to Telegram." Done.
Previously, memories were loose notes. Now, every memory follows a three-part template:
Claude no longer runs mkdir or checks if the memory directory exists—cuts unnecessary tool calls and prevents accidental setup steps.
Why it matters: Better context retention across sessions. Claude recalls not just what you told it, but why it matters and when to apply it. This is the foundation for coherent, stateful collaboration.
/v Command)Toggle it on, speak your instruction, Claude executes. Not a novelty. For developers who think faster than they type, or who want to dictate architectural decisions while reviewing code visually, voice mode is hands-free collaboration.
Claude for Excel and Claude for PowerPoint now share conversation context seamlessly. Pull data from spreadsheets, build tables, update a deck without re-explaining any step. Skills work in both add-ins. Team members can run a standard workflow (variance analysis, client deck) with a single click.
The revamped Skills Creator lets you write evals, benchmark performance, catch regressions, and verify trigger conditions—all without code.
Concrete example from the changelog: The PDF skill struggled with non-fillable forms. Using evals, the team isolated the failure, implemented a coordinate-anchoring fix, and verified consistent output. Multi-agent support enables parallel testing.
Ask Claude about Prompt Caching, Extended Thinking, Effort Controls, tools—directly in your terminal. No documentation digging. Instead of hunting through docs, you get actionable guidance on integrating API capabilities within your current project. Claude Code functions as a live API reference.
Every session now asks: Low, medium, high, or max? This controls how deeply Claude reasons, how long it works, and how much it costs. You can trigger "ultra" mode for complex multi-step problems requiring maximum reasoning budget. Clear effort levels from the start mean predictable results and efficient token usage.
A fleet of parallel agents scans your codebase (not just the diff) hunting for type mismatches, logic errors, architectural regressions. Catches bugs human reviewers miss. Costs $15-$25 per run. Available for Team and Enterprise subscribers.
This is the same system Anthropic runs on nearly every internal PR. It doesn't approve PRs but closes the review gap.
Build and view interactive charts, diagrams, visualizations directly in the conversation. Someone built an interactive instrument panel from a Cessna 172 directly in chat. Early, but the potential for education, data exploration, and architecture visualization is real.
If you write code daily: Start with /btw and /loop. They change the rhythm of how you work. Then explore Skills Creator evals.
If you build reports and presentations: Get Excel and PowerPoint add-ins. Shared context and one-click skills save hours of copy-paste.
If you manage a team: Desktop scheduled tasks with Telegram integration. Automated briefings and build alerts landing in a group chat.
If you build on the Claude API: The new API skill is your fastest path to answers about caching, thinking modes, tool integration.
12 Claude Code updates that shipped this week and what each one does
This article summarizes technical newsletters and curated links for developers. All views and opinions expressed here are for educational purposes. Verify claims and evaluate tools based on your specific needs before adopting them in production.