What's New in Claude Code — August 2026: Design in the Terminal, Shorter Answers, Permissions You Write as Sentences

What's New in Claude Code — August 2026: Design in the Terminal, Shorter Answers, Permissions You Write as Sentences

Anthropic sent Claude Code users a round-up this week of what shipped in August. The theme: the tool has outgrown "a coding assistant in your terminal" and become a wider workspace — interface design, response-style control, and execution permissions governed by rules written in plain language.

Here is what actually changed, and what it means for an engineering team or a product owner.

1) The design skill: idea to editable interface without leaving the terminal

/design is a new skill released as a research preview on 17 August 2026. Give it an idea, a screenshot, or an existing design, and it returns editable artboards in Claude Design — built on artifacts, and available in both the CLI and the desktop app.

The value is that the whole loop stays in one place: describe the screen, get several options, pick one, adjust it, then have Claude write the code that implements it. And because the output is a real design file rather than an image, it is shareable: on Team and Enterprise plans, anyone you make an editor can change it and publish a new version.

The smartest use comes before you write code, not after: ask for several options for a screen before you commit to one.

2) The Concise output style: result first, no preamble

A recurring complaint from developers is that Claude's responses run long. The answer is a built-in output style called Concise: it leads with the result, drops the preamble and the step-by-step narration, and keeps replies short by default. The engineering work itself is not shortened, and if you ask for a detailed explanation you get it in full.

One important exception: it never trims error reports, security warnings, or confirmations for destructive actions.

Turn it on via /config then Output style, or directly in settings:

{ "outputStyle": "Concise" }

It requires version 2.1.237 or later, and brings the built-in styles to five: Default, Proactive, Concise, Explanatory and Learning. A technical note: the output style is part of the system prompt, which is read once at session start, so a change takes effect after /clear or in a new session.

3) Auto-continue after a usage limit

Hitting a usage limit used to mean the session stopped and you waited for the reset before typing "continue." Now the session resumes from where it stopped as soon as the limit clears.

This does not raise your limits, but it removes the need to watch the clock. Its real value shows up in long tasks left running between sessions.

4) Auto mode is the default, and its rules are plain sentences

Since 14 August 2026, auto mode is the default permission mode for new sessions on Pro, Max and Team plans. In it, you do not review every action before it runs; a second model called the classifier reviews it and decides what passes and what stops. If you set your own default mode it stays in place unless you accept the switch prompt, an organization-managed default does not change, and you can pin your own with defaultMode in settings.

The practical part is that you steer that classifier with rules written as ordinary sentences rather than complex expressions. Entries under hard_deny block unconditionally, entries under soft_deny block unless a direct request or an allow entry overrides them, and adding $defaults keeps the built-in rules in force:

{
  "autoMode": {
    "soft_deny": [
      "$defaults",
      "Never run database migrations outside the migrations CLI"
    ],
    "hard_deny": [
      "$defaults",
      "Never send repository contents to third-party code-review APIs"
    ]
  }
}

These rules live in your user settings, so project-level settings cannot overwrite them and they stay in force across every project. That matters for any team that wants one security policy a single repository cannot break.

To sanity-check your rules before relying on them, claude auto-mode critique tells you which entries are ambiguous, redundant, or likely to cause false positives. Alongside it, claude auto-mode config prints the effective configuration, claude auto-mode defaults prints the built-in rules, and claude auto-mode reset restores them.

5) Release details: a default model, GitLab, and sessions that call each other

Beyond the four headlines, the recent releases carry small changes with outsized day-to-day impact:

• A default model for new sessions (2.1.236). The ANTHROPIC_DEFAULT_MODEL environment variable sets the model a session starts on when nothing else selects one; --model, ANTHROPIC_MODEL, a saved selection and an organization default all take priority.
• Auto-continue specifics (2.1.234). In the CLI it happens automatically and can be switched off in /config; in the desktop app you enable it on the session-limit card. The weekly-limit card does not offer it.
• GitLab merge request URLs in worktree (2.1.233). Pass a GitLab merge request URL to --worktree and Claude Code creates the branch from it, and on a branch with an open merge request a clickable badge appears in the footer (2.1.234). Marketplaces also clone bare gitlab.com URLs now (2.1.232). A meaningful change for any team on GitLab rather than GitHub.
• Mentioning other sessions (2.1.232). When another session is open on the same machine, type @ and its name in your prompt and Claude sends the message to that session. macOS and Linux only, and unsupported on Bedrock, Claude Platform on AWS, Google Cloud's Agent Platform and Foundry.
• Fork mode on by default (2.1.232). In interactive sessions Claude can spawn a fork subagent that starts with the full conversation instead of a blank context, and runs the subagents it spawns in the background. It is off by default with -p and in the Agent SDK, and enabled with CLAUDE_CODE_FORK_SUBAGENT=1.
• Code review in the background (2.1.232). At high, xhigh or max effort, /code-review runs as a background agent instead of taking over the session. The hosted review service remains Team and Enterprise only.
• Notable fixes (2.1.238). Remote Control no longer drops on a slow sign-in renewal or a brief rejection from a VPN or proxy, and messages sent from the web or desktop mid-turn stay in the transcript. Long sessions no longer grow in memory without limit, and a custom output style now holds for the whole session instead of drifting back to the default voice.

How an engineer inside Anthropic actually works

The newsletter included a striking quote from an engineer on the Claude Code team describing her daily setup: two lead agents that keep each other accountable and restart the other if either fails, delegating to tech lead or PM agents across the eight to ten projects she runs at any one time, with five to ten IC agents per project. Despite that scale she writes only thirty to fifty prompts a day, her IC agents work autonomously for two or three days at a stretch, and all of them communicate directly through the agent messaging tool.

The most telling number is not the agent count but how her time splits: sixty percent with the leads, thirty-five with a project lead, and five when something has gone off the rails. The leverage comes from a clear delegation structure, not from more tooling. Anyone chasing a similar result should start by defining who decides, who executes and who supervises, before spinning up one more agent.

Anthropic published two related posts the same week: one on running efficient sessions that get the most value from every token, and one setting out five operating principles drawn from interviews with more than a dozen fast-growing startups.

Other news in the same update

• The fifty percent increase to weekly limits has been extended through 31 August 2026 for Pro, Max, Team and seat-based Enterprise plans. Anthropic says it hopes to make this permanent, but demand for its models means it needs more time to work out how.
• The browser pane in the desktop app. Open it with a keyboard shortcut or from the Views menu, then click any element on the page and tell Claude what to change. Much faster than describing the element in words.

What this means for your team

First: the distance between design and implementation just shortened. /design does not replace your designer, but it removes the wait-for-the-mockup cycle in the early stages of internal features.

Second: making auto mode the default moves security responsibility from clicking approve to writing the policy. Anyone who has not written their rules is running on the built-ins alone — a decision that should be made deliberately, not by silence.

Third: the direction is clearly toward running several agents in parallel: fork on by default, sessions calling each other, code review running in the background. A team without a written delegation structure will get parallel chaos rather than parallel throughput.

Fourth: the release cadence is now weekly. Treating Claude Code as install-it-and-forget-it means missing changes that touch permissions and security directly, not just cosmetic features.

The Origami view

We follow these updates because they touch how we work day to day, not because they are tech news. GitLab support matters to us directly since our repositories live there, and a permission policy written into user settings solves an old problem: one security rule that a single project cannot break. The conclusion we apply to ourselves and recommend to our clients is the same: before adding agents or tools, write your policy and your delegation limits. Tools multiply what you already have, and if what you have is chaos, they multiply that.

Sources

• Claude Code team newsletter, 22 August 2026 — the original announcement.
• Official Claude Code documentation — output styles, permission modes, the desktop application, and the What's New page for week 32.
• Claude Code changelog for versions 2.1.232 through 2.1.238.
• Anthropic blog — the auto mode default announcement, plus the posts on session management and the startup guide.