Claude 4 Skills Improvements and New Features

Claude 4 brought meaningful improvements to the existing skills system If you’ve been using Claude Code skills for development workflows, here’s what changed and how to take advantage of it.

Enhanced PDF Processing

The pdf skill’s extraction engine improved significantly in Claude 4, particularly for multi-column layouts and scanned documents. Table detection is more reliable, which matters when pulling structured data from technical specifications and financial reports.

Batch processing capability was added in this version:

/pdf process all PDFs in invoices/ and extract the total, vendor name, and invoice date from each. Output as a CSV table.

Previously, you’d handle each document individually. The skill now handles the iteration internally, which reduces session length and token cost for document-heavy workflows.

For complex layouts that previously produced garbled extraction:

/pdf extract the data table on page 4 of multi-column-report.pdf. The table has 3 columns: Quarter, Revenue, YoY Growth.

Being explicit about table structure still improves accuracy, but the skill’s baseline performance on complex layouts is substantially better.

Smarter TDD Edge Case Detection

The tdd skill now identifies edge cases more proactively based on code patterns rather than requiring you to specify them.

/tdd write tests for this function and flag any edge cases I should handle: [paste function]

For a function accepting numeric limits, the skill now suggests boundary tests at -1, 0, 1, and max value without prompting. For async functions, it includes rejection paths. The test output is more complete out of the box.

Framework support expanded to include Vitest and Bun Test alongside pytest and Jest:

/tdd write Vitest tests for this TypeScript utility: [paste code]

/tdd write Bun Test tests for this module, using the describe/it/expect syntax: [paste code]

Fixture cleanup also improved. The skill now includes teardown logic in generated test suites, which previously had to be added manually.

Supermemory Semantic Search

The most significant supermemory skill improvement in Claude 4 is natural language retrieval. Previously, retrieval required exact key names. Now you can query by description:

/supermemory Why did we choose Redis over Memcached?

/supermemory What decisions did we make about the authentication flow?

The skill searches stored context semantically and surfaces relevant entries. For projects with months of accumulated decisions, this is substantially faster than remembering exact key names.

Team sharing is also new — shared context stores allow multiple developers to access the same project knowledge:

/supermemory store shared: onboarding-notes = monorepo uses pnpm, test with `pnpm test`, deploy with Railway CLI

New team members can recall the shared store immediately rather than waiting for documentation to be written.

Frontend Design Verification Depth

The frontend-design skill now checks more dimensions of design compliance, including layout shift metrics, accessibility violations, and responsive behavior across standard breakpoints.

/frontend-design audit this component at 320px, 768px, and 1440px widths. Report any layout shifts or overflow issues: [paste component]

/frontend-design verify this component against our design tokens — primary: #1A73E8, spacing: 8px grid, radius: 4px. List any violations: [paste component]

For teams enforcing a design system, the improved token verification catches violations that previously slipped through manual review.

Webapp Testing: Screenshots and Visual Regression

The webapp-testing skill added visual regression testing and video recording in Claude 4:

/webapp-testing take a baseline screenshot of the dashboard at http://localhost:3000/dashboard and save it as dashboard-baseline.png

/webapp-testing compare the current dashboard against dashboard-baseline.png and report any visual differences above 5% pixel change

/webapp-testing record a video of the checkout flow on http://localhost:3000 and save to checkout-flow.webm

The visual regression capability is the most practically useful addition for frontend teams — catching layout regressions without manual inspection.

Spreadsheet Charts and New Types

The xlsx skill gained waterfall charts, treemaps, and sparklines in Claude 4:

/xlsx create a waterfall chart in budget.xlsx showing starting balance, monthly inflows, monthly outflows, and ending balance for Q1

/xlsx add sparklines to the summary row in metrics.xlsx showing the 12-month trend for each KPI

These chart types were previously difficult to generate programmatically and required manual steps in Excel. The skill handles the chart object configuration directly.

Skill Auto-Invocation Triggers and Hook Configuration

Claude 4 introduced configurable auto-invocation triggers. Skills can now activate based on file patterns you define in ~/.claude/settings.json under the "hooks" key.

The hook types available are:

PreToolUse — fires before Claude executes a tool (Bash, Write, Edit, etc.)
PostToolUse — fires after tool execution with access to the output
Notification — fires when Claude sends a notification

Hooks receive context about the current operation as environment variables. CLAUDE_TOOL_NAME, CLAUDE_TOOL_INPUT, and CLAUDE_TOOL_OUTPUT_PATH are available depending on the hook type.

A minimal example that logs every bash command and auto-formats file writes:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": { "tool_name": "Bash" },
        "command": "echo 'Running bash command' >> ~/.claude/audit.log"
      }
    ],
    "PostToolUse": [
      {
        "matcher": { "tool_name": "Write" },
        "command": "npx prettier --write \"$CLAUDE_TOOL_OUTPUT_PATH\""
      }
    ]
  }
}

You can also block dangerous operations by exiting with a non-zero code in a PreToolUse hook:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": { "tool_name": "Bash" },
        "command": "if echo \"$CLAUDE_TOOL_INPUT\" | grep -q 'rm -rf'; then echo 'Blocked: rm -rf not allowed' >&2; exit 1; fi"
      }
    ]
  }
}

The /tdd, /pdf, and other skills still work exactly the same way when invoked manually with /skill-name. Auto-invocation via hooks is an optional pattern for teams that want skills to surface contextually without manual invocation.

Getting Started with These Features

All improvements activate automatically when you use the updated skills — no configuration changes required for the core improvements. If you’re on an older Claude Code version, update first:

npm update -g @anthropic-ai/claude-code

Then verify you have Claude 4 by checking the model in your settings or running claude --version.

Best Claude Skills for Developers in 2026 — Top skills every developer should know
Claude Skills vs Prompts: Which Is Better? — Decide when skills beat plain prompts
Claude Skills Auto Invocation: How It Works — How skills activate automatically

*Built by theluckystrike — More at zovo.one *