Hey!

Two big features shipped in beta this week. Projects are a new way to let Claude coordinate work across multiple sessions for you from one conversation, and Claude Design, Slides and the new Docs now work right in the desktop app and on web.

Projects in Claude Code (beta)

The Projects you may know from the Claude app got a full redesign this week, and the new version is in beta in Claude Code first for select users.

Until now a project was a folder for your chats, files and instructions. The new version actually runs the work.

A project in the Claude Code desktop app: the conversation on the left where Claude reports what each thread did, and the Threads pane on the right grouping its threads into Working, Idle and Resolved

Instead of managing many sessions yourself, you tell Claude what needs to get done in one conversation and it splits the work into threads, runs each as a cloud session on its own branch, and keeps going after you close your laptop.

It's quickly become my favorite way to work with Claude Code. I stay focused on my own work and just check the project to see what got done and where it needs me!

Rolling out now to Pro and Max on cloud sessions, with a waitlist if you don't see it yet.

Claude Design, Slides and Docs in Claude Code on desktop and web (beta)

Claude Code on desktop making a Claude Design canvas from the conversation, with the architecture board open in the preview pane next to it

Claude Design, Claude Slides and the new Claude Docs now work inside Claude Code on desktop and web. Claude can make a UI mockup, a review deck or an RFC from your repo in the same session, and you can edit the result right in the preview pane.

I switch between apps even less now, most days it's just Slack and Claude. All three are in beta on paid plans.

Try it in Claude | Learn more about Claude Design

Other news

Fable 5.1 Build Days: the Claude community is hosting buildathons in cities around the world until September 25. Bring a problem, an idea, or just show up, and find one near you on the community events page.

Get more out of Claude Code

Eval your plugins and skills

Claude Code can now test whether your plugin or skill actually improves Claude's answers. Run claude plugin eval init in your plugin's folder, tell Claude what a good result looks like, and it writes the test cases and graders for you.

Then, run claude plugin eval .:

Running claude plugin eval . in a terminal prints a table of cases with their scores with and without the plugin, the difference, the number of runs and the cost

Claude runs each case with and without your plugin and scores both side by side, but each case is a real model call, so keep an eye on usage (requires v2.1.269+).

Keep your computer awake

If you leave a long session running on your laptop, the desktop app can now stop your machine from sleeping for just that session.

The session menu in the Claude Code desktop app with the Keep computer awake toggle, labelled Only for this session, switched on

Turn on Keep computer awake in that session's ⋮ menu and your machine stays up for that session, including between turns.

Get the desktop app →

More control over subagents

You get a lot of control over a custom subagent from its frontmatter. You can pick its model and effort, limit its tools, cap its turns, give it its own worktree or memory, and you can now also skip your CLAUDE.md files with omitClaudeMd: true.

.claude/agents/log-scanner.md

---

name: log-scanner

description: Scans CI logs for the first failing step and reports it

model: sonnet

effort: low

memory: project

omitClaudeMd: true

---

Find the first failing step in the log you are given and report the command, the error and the file it points to.

Every subagent normally loads your CLAUDE.md files into its fresh context first, but a narrow worker like this log scanner doesn't need your repo conventions. Leaving them out keeps its context small and each spawn quick to start.

Read more about subagent frontmatter →

How to Claude like Anthropic

"I've wondered if a lot of time is spent working around issues that I don't notice since I don't supervise my sessions as much as I used to. (The answer is yes.) I've added a papercuts section to my CLAUDE.md: Maintain ~/code/papercuts.md, a global log shared by all my Claude sessions of anything that slowed down development. When you lose time to one mid-session, append date · symptom · fix · project. Check this file first when tooling fails mysteriously. It's already found a lot of recurring issues."

– Pedram Navid, Engineer at Anthropic

"After a long session I want to see the decisions Claude made along the way, not re-read the whole thread. So I ask: In your own words, without jargon, narrate each step in the order it actually happens, and bold inline anything in the current PR that has drifted from my most recently stated intent. It's a great way to find where my mental model differs from Claude's. Then I paste the response in a quote block and respond to the parts that stand out."

– Sam, Engineer at Anthropic

Other highlights

ADDED

2.1.269

/output-style [name] to list and switch output styles: Run /output-style to list styles, or add a style name to switch mid-session. Also works over Remote Control and in cloud sessions.

2.1.273

A notice when an MCP server stays disconnected: When a remote MCP server drops and reconnection gives up after five attempts, Claude Code now tells you and points you to /mcp to retry.

2.1.271

Per-command network hosts in auto mode with sandboxing: With sandboxing on in auto mode, Claude lists the network hosts a Bash command needs and the classifier reviews them with the command. Approved hosts open for that one command only.

2.1.269

A diff of the files a Bash command changed, with bashEditDiffEnabled: When a Bash command changes files in a Git repo, you now see the diff right after. On by default in auto mode; set "bashEditDiffEnabled": true to get it in every permission mode.

CHANGED

2.1.268

/plugin changes now apply when you close the menu: Plugin changes in /plugin now take effect when you close the menu, since Claude Code runs /reload-plugins for you. If that would invalidate the prompt cache, it warns and waits for /reload-plugins --force.

2.1.271

The spinner now names the hook it is waiting on: While a SessionStart, UserPromptSubmit, PreToolUse or SessionEnd hook runs, the spinner names it with the elapsed time. Esc cancels a prompt waiting on SessionStart.

2.1.271

Workflows now pause at your usage limit instead of dropping agents: A dynamic workflow that hits your usage limit now pauses and resumes when the limit resets instead of dropping agents.

FIXED

2.1.270

Read-only git commands asking for permission: In 2.1.269, read-only git commands could start prompting for permission in long sessions. Fixed in 2.1.270.

2.1.268

High CPU in long idle sessions: Long idle sessions no longer pin a CPU core, and rapid terminal focus reports during a recap no longer keep it high.

2.1.269

/goal runs stalling silently: After an API error or a dropped connection, a /goal run now retries on its own, then pauses and says why. If you hit your usage limit, it pauses and resumes when the limit resets.

That's it for this week! To see all that shipped, take a look at our full changelog. If you hit a bug, run /feedback in Claude Code and it comes straight to us.

Thanks for building with Claude Code,

Lydia and the Claude Code team