Week 14: Computer Use in CLI, /powerup Lessons, and 14 New Videos

Claude can now drive native desktop apps, click through UI, test its own changes, and fix what breaks — all from your terminal. Not just browser flows anymore; works with simulators, desktop apps, and GUI-only workflows.

// Just ask Claude Code to interact with desktop apps:
// "Open the Figma desktop app and export the selected frame as PNG"
// "Test the iOS simulator and verify the login flow works"
// "Click the Submit button in the app and check the response"

Create a custom skill that adds a visual status line showing context window usage percentage. Prevents unexpected auto-compaction that disrupts workflows. From Aakash Gupta's advanced setup.

// Ask Claude Code to create this skill:
// "Create a skill that shows my current context window usage
// as a percentage in the terminal status line.
// Warn me when I'm above 70% so I can compact or delegate."

Persistent memory remains a core unmet need. Two community approaches: (1) Obsidian + session hooks + daily reflection cron jobs for full architecture, or (2) NotebookLM as a lightweight RAG retrieval layer. Both reduce repeated token costs.

Built-in interactive lessons that teach you Claude Code features step by step. Anthropic embedded feature discovery directly inside the product so you don't miss fast-moving releases.

/powerup

Browse release notes interactively inside Claude Code. Pick any version to see what changed. Anthropic is improving discoverability of their fast-moving releases.

/release-notes

This week's discourse was heavily leak-driven. Many videos mixed real architecture observations with speculation. Do NOT treat leaked internals as stable product features. Avoid downloading or running leaked code due to supply-chain attack risks.

Environment variable that eliminates terminal flickering during long sessions with heavy scrollback. Set it before launching Claude Code for smoother output.

# Add to your shell profile (.bashrc, .zshrc, etc.):
export CLAUDE_CODE_NO_FLICKER=1

# Or set it inline:
CLAUDE_CODE_NO_FLICKER=1 claude

For managed environments: Claude Code now blocks startup if fresh remote settings cannot be fetched. Stronger administrative guarantees for enterprise deployments.

// In managed settings:
{
  "forceRemoteSettingsRefresh": true
}

MCP tool authors can now set _meta["anthropic/maxResultSizeChars"] to keep large but valid outputs inline instead of spilling to disk. Important for schemas, file trees, and similar payloads.

// In your MCP tool response, add _meta:
{
  "_meta": {
    "anthropic/maxResultSizeChars": 50000
  },
  "result": "...large output..."
}

Plugins can now ship executable binaries in a bin/ directory that are automatically added to PATH. Plugins are becoming packaged capability bundles, not just prompts.

Large marketplace-oriented skill distribution. Exposes official-style commands like /plugin marketplace add and bundle installs by domain. One of the clearest examples of a skill catalog model.

// Install via Claude Code:
/plugin marketplace add alirezarezvani/claude-skills

// Or clone manually:
git clone https://github.com/alirezarezvani/claude-skills.git ~/.claude/commands/claude-skills

Hooks can now listen for PermissionDenied events and use the 'defer' decision for more expressive workflow control. Useful for headless or semi-automated sessions.

// In settings.json hooks:
{
  "hooks": {
    "PermissionDenied": [{
      "command": "echo 'Permission denied for: $CLAUDE_TOOL_NAME' >> /tmp/denied.log",
      "decision": "defer"
    }]
  }
}

Directly addresses Claude Code failure modes through hooks + CLAUDE.md. Covers post-edit verification, stop verification, truncation warnings, and destructive-command blocking. Most actionable hardening repo this week.

// Clone and install:
git clone https://github.com/iamfakeguru/claude-md.git
cd claude-md && ./install.sh

// What it adds:
// - Post-edit verification hooks
// - Stop verification before marking done
// - Truncation warnings
// - Destructive-command blocking

New setting that prevents skills from executing shell commands. Anthropic acknowledges that skills and plugin commands can be a security boundary, not only a convenience layer.

// In settings.json:
{
  "disableSkillShellExecution": true
}

Security skills from Trail of Bits (a respected security firm). Includes Semgrep, supply-chain review, differential review, and mutation testing skills. Serious security practitioners investing in Claude Code skill packaging.

// Install via marketplace:
/plugin marketplace add trailofbits/skills

// Or clone:
git clone https://github.com/trailofbits/skills.git ~/.claude/commands/trailofbits-skills

The /cost command now shows per-model usage breakdown and cache-hit rates. Better visibility into where your tokens are going, which supports smarter model-routing and cost hygiene.

/cost

Treat Claude Code as the agentic shell and swap the model engine. Use Ollama (local) or OpenRouter (cloud) for boilerplate, summarization, and simple tasks. Reserve premium models for complex reasoning.

// Local with Ollama:
// 1. Install Ollama and pull a model:
ollama pull qwen3.5:9b

// 2. In settings.local.json, override:
{
  "ANTHROPIC_BASE_URL": "http://localhost:11434/v1",
  "ANTHROPIC_AUTH_TOKEN": "ollama"
}

// Cloud with OpenRouter:
// In settings.local.json:
{
  "ANTHROPIC_BASE_URL": "https://openrouter.ai/api/v1",
  "ANTHROPIC_AUTH_TOKEN": "your-openrouter-key"
}

Demonstrates AI-assisted reverse engineering of the leaked Claude Code source. Interesting as a software-modification technique demo, but based on leaked proprietary code. Watch for the workflow, not as a model to follow.

Best live walkthrough of recent official features: /voice, /loop, Remote Control, Auto Memory, and browser/computer-use. Translates release notes into lived constraints. High-signal official-overlap tutorial.

Argues terminal-only usage is the wrong abstraction for business users. Introduces a 'Command Centre' web UI concept and 'Agentic OS' framework using persistent markdown files for brand context, skills, and memory.

Strongest general-purpose onboarding video this week. Covers editor integration, CLAUDE.md as persistent memory, plugin installation (Playwright, Superpowers, Context7), GitHub MCP workaround, and slash commands.

In-depth leak analysis revealing internal architecture, unreleased features (Coordinator Mode, Ultraplan, Kairos), and prompt-engineering advice from internal guidelines. Useful for prompt tips, but leak-dependent.

Demonstrates routing Claude Code through Ollama (local) or OpenRouter (cloud) for cost savings. Treats Claude Code as the agentic shell and the model as a replaceable engine. Best for low-stakes, high-volume tasks.

Power-user checklist from leak observations: /init, /plan, /compact, /cost, CLAUDE.md centrality, and settings-level control. Good conceptual accelerator but leak-derived, not authoritative documentation.

OpenRouter rerouting tutorial for using Claude Code with free/cheaper models. The technical core is real but wrapped in marketing hype. Extract the JSON-editing and routing pattern, not the sales pitch.

Strong commentary on Anthropic's DMCA response, supply-chain risks from leaked code, and engineering practices. More of a software-process lesson than a Claude Code tutorial. Useful for engineering leaders.

Most balanced leak summary: architecture notes (Bun, multi-agent, anti-distillation, profanity logging, compaction), legal/copyright questions, and sober tone. Good context without sensationalism.

Most novel third-party integration demo this week. Uses MCP bridge to read live TradingView DOM data instead of screenshots. Impressive but fragile: depends on debugging port, DOM structure, and carries financial risk.

One of the best power-user workflow videos this week. Treats Claude Code as an OS: context window management, sub-agent delegation, hooks, project isolation, global .claude rules. Operational design, not just feature discovery.

Robust architecture for a custom second-brain agent: Claude Code + Obsidian + session hooks + cron jobs + secure APIs + modular memory files. Security-conscious approach. Partly promotional but architecturally strong.

Lightweight external memory strategy using NotebookLM as a retrieval layer. Summarize and store context externally, query when needed. Practical and accessible, though title overpromises. Good entry point for second-brain patterns.

Workflow orchestration for multi-session control, role-based execution, and multi-CLI coordination. Powerful but heavy — most users won't need this much infrastructure. Best for complex multi-agent setups.

// Install globally:
npm install -g claude-code-workflow

// Quick start:
ccw install -m Global

// Then use ccw commands to manage workflows

Instead of doing web research or multi-step tasks in your main session, delegate to a sub-agent. This preserves your main session's context window for the actual work. Key pattern from power users this week.

// Tell Claude Code:
// "Research the best React form libraries and write a summary.
// Do this in a sub-agent so we don't use up our main context."