Projects in Claude Code, and Claude Design, Slides and Docs on desktop and web

Anthropic's weekly Claude Code newsletter of September 19, 2026 leads with two beta features. Projects lets Claude coordinate work across multiple sessions from one conversation, and Claude Design, Claude Slides and the new Claude Docs now run inside Claude Code on desktop and on the web. Both are beta, and both extend the same direction Anthropic took when it merged Cowork and chat into one Claude three days earlier: fewer places to choose between, and work that keeps running while you are away.

Projects now run the work instead of storing it. Until this week a project was a folder for your chats, files and instructions, the way you know it from the Claude app. The redesigned version does something different: you describe in one conversation what needs to get done, and Claude splits the work into threads, runs each thread as a cloud session on its own git branch, and keeps going after you close your laptop. The project view shows the conversation on one side and a Threads pane on the other, grouping threads into Working, Idle and Resolved, so you check in on what got done and where a thread is waiting on you rather than managing a stack of sessions yourself. It is rolling out now to Pro and Max on cloud sessions, in Claude Code first and for selected users, with a waitlist if you do not see it yet.

Design, Slides and Docs inside Claude Code. Claude Design, Claude Slides and the new Claude Docs now work in Claude Code on desktop and on the web. In the same session where Claude reads your repository, you can ask for a UI mockup, a deck for a design review, or an RFC drawn from the code itself, and edit the result directly in the preview pane next to the conversation. All three are in beta on paid plans. For anyone who already works with Docs, Slides and Design in the merged Claude, this is the same set of surfaces, now also available where the code lives.

You can now measure whether a plugin or skill actually helps. Claude Code can test whether your plugin or skill improves Claude's answers. Run claude plugin eval init in the plugin's folder, describe what a good result looks like, and Claude writes the test cases and graders for you. Running claude plugin eval . then executes each case twice, once with your plugin and once without, and prints a table with both scores, the difference, the number of runs and the cost. Two caveats worth taking seriously: every case is a real model call, so a broad suite costs real tokens, and the feature requires version 2.1.269 or later. This is the first built-in answer to a question most skill authors could previously only guess at: does this skill make the output better, or does it merely feel that way?

Keep computer awake. A long session on a laptop that goes to sleep is a session that stops. The desktop app now has a Keep computer awake toggle in a session's menu, scoped to that one session, and it holds between turns as well as during them. It is a small fix for a failure mode that costs a full run when it hits.

More control over subagents. A custom subagent is configured from the frontmatter of its file in .claude/agents/. You could already pick its model and effort level, limit which tools it may use, cap the number of turns, give it its own worktree or its own memory. New this week: omitClaudeMd: true, which stops the subagent from loading your CLAUDE.md files. Every subagent normally reads those into its fresh context first. For a narrow worker, say an agent that scans a CI log for the first failing step, your repository conventions are dead weight: leaving them out keeps the context small and the spawn quick.

Smaller changes worth knowing. /output-style now lists your output styles, or switches to one mid-session if you add a name, and it works over Remote Control and in cloud sessions (2.1.269). When a remote MCP server drops and reconnection gives up after five attempts, Claude Code now says so and points you at /mcp instead of failing quietly (2.1.273). With sandboxing on in auto mode, Claude lists the network hosts a Bash command needs and the classifier reviews them along with the command; approved hosts open for that one command only (2.1.271). When a Bash command changes files in a git repository you now see the diff right after, on by default in auto mode and available everywhere else with the bashEditDiffEnabled setting (2.1.269). Plugin changes in /plugin apply when you close the menu because Claude Code runs /reload-plugins for you, warning first if that would invalidate the prompt cache (2.1.268). The spinner now names the hook it is waiting on, with elapsed time (2.1.271). And a dynamic workflow that hits your usage limit now pauses and resumes when the limit resets, instead of dropping agents (2.1.271).

Fixes. Read-only git commands that started prompting for permission in long sessions in 2.1.269 were fixed in 2.1.270. Long idle sessions no longer pin a CPU core (2.1.268). And a /goal run that hit an API error or a dropped connection used to stall silently; it now retries on its own, then pauses and says why, and pauses and resumes around a usage limit (2.1.269).

Two working habits from Anthropic's own engineers. The first is a papercuts log. Pedram Navid keeps a global file that all his Claude sessions share, with a line in his CLAUDE.md instructing Claude to append date, symptom, fix and project whenever a session loses time to something, and to check that file first when tooling fails mysteriously. His observation is the useful part: the less you supervise sessions, the more time goes into working around issues you never see. The second is a review prompt from Sam, an engineer at Anthropic, for the end of a long session: ask Claude to narrate each step in the order it actually happened, in plain language, and to bold anything in the current pull request that has drifted from your most recently stated intent. It surfaces where your mental model and Claude's have come apart, without rereading the whole thread.

Also mentioned. The Claude community is running Fable 5.1 Build Days, buildathons in cities around the world, until September 25. Anthropic points to its community events page to find one nearby.

What this means for readers of this site. Nothing here changes how Cowork works, but the direction is worth noting. Projects is the same idea as the merged Claude, applied to code: you hand over an outcome instead of managing the sessions that produce it, and the work continues while you are not watching. Docs, Slides and Design arriving in Claude Code means those surfaces are now available wherever you happen to be working, which is the point of merging them into one Claude in the first place. And plugin evals matter to anyone maintaining skills: the same question, does this skill actually improve the answer, applies to the skills you write for Cowork, even though the eval command itself lives in Claude Code for now.