10 Skills of the Day

Set package cooldown to 72 hours across all your package managers. pnpm: resolution-time=72h, uv: --exclude-newer, npm via .npmrc. This single config change would have protected you from the LiteLLM attack. Willison's survey covers all seven managers.

Install Lasso Security's claude-hooks as a PostToolUse defender today. 50+ regex patterns catch prompt injection in tool results before Claude processes them. Takes 5 minutes to set up. If you run --dangerously-skip-permissions without this, you're running undefended against documented attacks.

Add CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 to your shell profile. New in Claude Code v2.1.83. Strips Anthropic keys, AWS credentials, database creds, and API tokens from every subprocess environment. Not enabled by default. One line in .zshrc blocks environment variable exfiltration.

Use claude -w (worktree flag) to run parallel Claude Code sessions on separate branches. Each session gets full filesystem independence via git worktrees. Review completed work in one while the other is still coding. incident.io runs 4-5 sessions as their default workflow.

Front-load deterministic context before agent loops, Stripe-style. Don't let your agent discover what it needs. Gather file paths, API docs, and relevant code before the LLM runs. The pattern: deterministic prefetch, then curate 10-15 tools from your catalog, then run the agentic loop. Saves tokens and produces better results.

Write a DESIGN.md file for your projects using Google Stitch's export. Extract your design system (colors, typography, spacing) as a markdown file that coding agents read alongside CLAUDE.md. Place both in your project root. Now your agent knows architecture and design constraints simultaneously.

Audit your MCP server stack against the OWASP MCP Top 10. Check for hardcoded credentials, missing authentication, unvalidated inputs, and overly permissive tool scopes. The checklist has concrete remediation for each category. At 92% exploit probability with 10 plugins, every MCP server you remove reduces your attack surface exponentially.

Use Playwright CLI instead of Playwright MCP for agent-driven browser testing. CLI uses 27K tokens per task versus 114K with MCP. A 4x reduction. CLI saves screenshots to disk and lets agents selectively read what's needed rather than streaming accessibility trees into context. Install via npm i @playwright/cli.

Curate agent tool catalogs, don't dump them. Stripe curates ~15 tools from 400+ per run. LangChain's data shows accuracy jumps from 17% to 92% with progressive skill disclosure. Build a tool catalog, tag each by use case, and dynamically select the relevant subset per task. More tools equals worse performance.

Patch n8n to 2.10.1 immediately and disable all public Form nodes until patched. Four CVEs at CVSS 9.4-9.5, including unauthenticated RCE via public forms. CISA deadline is today. 71,537 instances exposed. If you're running n8n with public-facing forms, this is the highest-priority action item in today's newsletter.