🧭 Claude Academy: Anthropic's 4D AI Fluency Framework Now Public

Anthropic has expanded Claude Academy — its free learning platform at anthropic.skilljar.com — with a cohesive curriculum built around the 4D AI Fluency Framework that Anthropic uses to onboard every new employee. The framework is now central to all entry-level courses and is freely available to individuals, teams, and educators worldwide.

The four Ds

• Delegation — deciding which tasks are genuinely suited to AI assistance and which require human judgement; the skill of task decomposition and handoff.
• Description — writing precise, unambiguous prompts: providing context, examples, constraints, and success criteria so Claude can act on intent, not inference.
• Discernment — evaluating AI outputs critically: identifying hallucinations, spotting reasoning gaps, and knowing when to distrust a confident-sounding answer.
• Diligence — maintaining appropriate human oversight; verifying important outputs independently and applying responsible-use practices in production workflows.

Three learning tracks

Claude Academy organises its 20+ self-paced courses into three distinct tracks:

• AI Fluency (no-code) — for professionals who want to work better with Claude without learning to program. Covers the 4D Framework, task delegation strategies, and department-specific use cases (marketing, legal, HR, finance).
• Developer — API integration, Claude Code, MCP server development, tool use, and agentic workflow design. Includes hands-on coding labs and sandbox environments.
• Cloud integration — deploying Claude via Amazon Bedrock and Google Vertex AI, including IAM configuration, throughput management, and fine-tuning with Constitutional AI data.

🧭 Self-Hosted Environments for Claude Code Now in Public Beta

Anthropic has opened the public beta of self-hosted environments for Claude Code, letting engineering teams on Team and Enterprise plans run Claude Code sessions entirely on infrastructure they operate. Instead of compute routing through Anthropic's cloud, Claude Code spins up on runners you provision — inside your VPC, behind your firewall, with access to your internal network.

What self-hosted environments enable

• Internal network access — Claude can reach internal databases, private APIs, documentation servers, and artifact registries without VPN tunnelling or IP allowlisting on the Anthropic side.
• Custom tooling — bring your own runner container. Pre-install compilers, linters, proprietary SDKs, or internal CLI tools; Claude Code sees them as if running on a developer's local machine.
• Compliance controls — all compute and intermediate data stay on your infrastructure. Model inference calls still go to Anthropic's API, but tool execution, file reads/writes, and terminal sessions run locally.
• Two deployment modes: fixed runners (always-on, low cold-start latency) and on-demand runners (spun up per session, lower cost at low volume). Both modes support horizontal scaling under Claude Code's existing session concurrency limits.

Configuration in a nutshell

# In your organisation's Claude Code settings
{
  "environment": {
    "type": "self-hosted",
    "runner": "on-demand",
    "image": "registry.internal/claude-runner:latest",
    "network": "vpc-prod"
  }
}

🧭 Claude Code Patch: Readline Keybindings, MCP Hardening, and Faster Startup

Today's Claude Code update is a targeted quality-of-life and security release. No headline features — just fixes and settings that make daily terminal work noticeably smoother.

New: keybindingFlavor setting

Set "keybindingFlavor": "readline" in your settings.json to switch Ctrl+W from "delete entire prompt" to "delete back to previous whitespace" — matching the behaviour you'd expect from Bash, Zsh, or any readline-based shell. For developers who live in the terminal, this is the fix to ship first.

// ~/.claude/settings.json
{
  "keybindingFlavor": "readline"
}

MCP plugin security hardening

MCP servers can no longer request file-system tool permissions beyond the scope declared in their manifest. Previously, a malicious or misconfigured MCP plugin could request broader Bash access at runtime by returning unexpected tool definitions. The new enforcement layer validates permissions at load time against the server's declared capability set and rejects unapproved escalations with a user-visible warning.

Other fixes

• Startup speed ~30% faster on repositories with large CLAUDE.md files — non-critical sections are now loaded lazily after the first prompt is ready.
• Memory leak patched in long-running sessions with many sequential tool calls; sessions now stay stable over hours rather than gradually ballooning.
• Cross-session messaging is more resilient — pending messages are buffered during network flaps and delivered on reconnect without loss.
• Terminal rendering glitch fixed for users on tmux with TERM=screen-256color where some ANSI escape codes were rendered as literal text.