Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
239 changes: 239 additions & 0 deletions skills/webperf-core-web-vitals/SKILL.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,239 @@
---
name: webperf-core-web-vitals
description: Intelligent Core Web Vitals analysis with automated workflows and decision trees. Measures LCP, CLS, INP with guided debugging that automatically determines follow-up analysis based on results. Includes workflows for LCP deep dive (5 phases), CLS investigation (loading vs interaction), INP debugging (latency breakdown + attribution), and cross-skill integration with loading, interaction, and media skills. Use when the user asks about Core Web Vitals, LCP optimization, layout shifts, or interaction responsiveness.
context: fork
---

# WebPerf: Core Web Vitals

JavaScript snippets for measuring web performance in Chrome DevTools. Execute with `evaluate_script`, capture output with `get_console_message`.

## Scripts

- `scripts/CLS.js` — Cumulative Layout Shift (CLS)
- `scripts/INP.js` — Interaction to Next Paint (INP)
- `scripts/LCP-Image-Entropy.js` — LCP Image Entropy
- `scripts/LCP-Sub-Parts.js` — LCP Sub-Parts
- `scripts/LCP-Trail.js` — LCP Trail
- `scripts/LCP-Video-Candidate.js` — LCP Video Candidate
- `scripts/LCP.js` — Largest Contentful Paint (LCP)

Descriptions, thresholds, and return schemas: `references/snippets.md`, `references/schema.md`

## Script Execution Patterns

Scripts fall into two execution patterns:

### Synchronous (LCP, CLS, LCP-Sub-Parts, LCP-Trail, LCP-Image-Entropy, LCP-Video-Candidate)

Run via `evaluate_script` and return structured JSON immediately from buffered performance data. The page must have already loaded.

### Tracking (INP)

INP requires real user interactions to measure. The workflow is:

1. Run `INP.js` via `evaluate_script` → returns `{ status: "tracking", getDataFn: "getINP" }`
2. **Tell the user:** "INP tracking is now active. Please interact with the page — click buttons, open menus, fill form fields — then let me know when you're done."
3. Wait for the user to confirm they've interacted.
4. Call `evaluate_script("getINP()")` to collect results.
5. If `getINP()` returns `status: "error"` → the user has not interacted yet. Remind them and wait.
6. For a full breakdown of all interactions, call `evaluate_script("getINPDetails()")` — returns all recorded interactions sorted by duration.

> The agent cannot interact with the page on behalf of the user for INP measurement. Real user interactions are required.

## Common Workflows

### Complete Core Web Vitals Audit

When the user asks for a comprehensive Core Web Vitals analysis or "audit CWV":

1. **LCP.js** - Measure Largest Contentful Paint
2. **CLS.js** - Measure Cumulative Layout Shift
3. **INP.js** - Measure Interaction to Next Paint
4. **LCP-Sub-Parts.js** - Break down LCP timing phases
5. **LCP-Trail.js** - Track LCP candidate evolution

### LCP Deep Dive

When LCP is slow or the user asks "debug LCP" or "why is LCP slow":

1. **LCP.js** - Establish baseline LCP value
2. **LCP-Sub-Parts.js** - Break down into TTFB, resource load, render delay
3. **LCP-Trail.js** - Identify all LCP candidates and changes
4. **LCP-Image-Entropy.js** - Check if LCP image has visual complexity issues
5. **LCP-Video-Candidate.js** - Detect if LCP is a video (poster or video element)

### CLS Investigation

When layout shifts are detected or the user asks "debug CLS" or "layout shift issues":

1. **CLS.js** - Measure overall CLS score
2. **Layout-Shift-Loading-and-Interaction.js** (from `webperf-interaction` skill) - Separate loading vs interaction shifts
3. Cross-reference with **webperf-loading** skill:
- Find-Above-The-Fold-Lazy-Loaded-Images.js (lazy images causing shifts)
- Fonts-Preloaded-Loaded-and-used-above-the-fold.js (font swap causing shifts)

### INP Debugging

When interactions feel slow or the user asks "debug INP" or "slow interactions":

1. **INP.js** - Start tracking. Tell the user to interact with the page and confirm when done.
2. Call `getINP()` to collect results once the user confirms.
3. Call `getINPDetails()` to see all interactions ranked by duration.
4. **Interactions.js** (from `webperf-interaction` skill) - List all interactions with timing
5. **Input-Latency-Breakdown.js** (from `webperf-interaction` skill) - Break down input delay, processing, presentation
6. **Long-Animation-Frames.js** (from `webperf-interaction` skill) - Identify blocking animation frames
7. **Long-Animation-Frames-Script-Attribution.js** (from `webperf-interaction` skill) - Find scripts causing delays

### Video as LCP Investigation

When LCP is a video element (detected by LCP-Video-Candidate.js):

1. **LCP-Video-Candidate.js** - Identify video as LCP candidate
2. **Video-Element-Audit.js** (from Media skill) - Audit video loading strategy
3. **LCP-Sub-Parts.js** - Analyze video loading phases
4. Cross-reference with **webperf-loading** skill:
- Resource-Hints-Validation.js (check for video preload)
- Priority-Hints-Audit.js (check for fetchpriority on video)

### Image as LCP Investigation

When LCP is an image (most common case):

1. **LCP.js** - Measure LCP timing
2. **LCP-Sub-Parts.js** - Break down timing phases
3. **LCP-Image-Entropy.js** - Analyze image complexity
4. Cross-reference with **webperf-media** skill:
- Image-Element-Audit.js (check format, dimensions, lazy loading)
5. Cross-reference with **webperf-loading** skill:
- Find-Above-The-Fold-Lazy-Loaded-Images.js (check if incorrectly lazy)
- Priority-Hints-Audit.js (check for fetchpriority="high")
- Resource-Hints-Validation.js (check for preload)

## Decision Tree

Use this decision tree to automatically run follow-up snippets based on results:

### After LCP.js

- **If LCP > 2.5s** → Run **LCP-Sub-Parts.js** to diagnose which phase is slow
- **If LCP > 4.0s (poor)** → Run full LCP deep dive workflow (5 snippets)
- **If LCP candidate is an image** → Run **LCP-Image-Entropy.js** and **webperf-media:Image-Element-Audit.js**
- **If LCP candidate is a video** → Run **LCP-Video-Candidate.js** and **webperf-media:Video-Element-Audit.js**
- **Always run** → **LCP-Trail.js** to understand candidate evolution

### After LCP-Sub-Parts.js

- **If TTFB phase > 600ms** → Switch to **webperf-loading** skill and run TTFB-Sub-Parts.js
- **If Resource Load Time > 1500ms** → Run:
1. **webperf-loading:Resource-Hints-Validation.js** (check for preload/preconnect)
2. **webperf-loading:Priority-Hints-Audit.js** (check fetchpriority)
3. **webperf-loading:Find-render-blocking-resources.js** (competing resources)
- **If Render Delay > 200ms** → Run:
1. **webperf-loading:Find-render-blocking-resources.js** (blocking CSS/JS)
2. **webperf-loading:Script-Loading.js** (parser-blocking scripts)
3. **webperf-interaction:Long-Animation-Frames.js** (main thread blocking)

### After LCP-Trail.js

- **If many LCP candidate changes (>3)** → This causes visual instability, investigate:
1. **webperf-loading:Find-Above-The-Fold-Lazy-Loaded-Images.js** (late-loading images)
2. **webperf-loading:Fonts-Preloaded-Loaded-and-used-above-the-fold.js** (font swaps)
3. **CLS.js** (layout shifts contributing to LCP changes)
- **If final LCP candidate appears late** → Run **webperf-loading:Resource-Hints-Validation.js**
- **If early candidate was replaced** → Understand why initial content was pushed down (likely CLS issue)

### After LCP-Image-Entropy.js

- **If entropy is very high** → Image is visually complex, recommend:
- Modern formats (WebP, AVIF)
- Appropriate compression
- Potentially a placeholder strategy
- **If entropy is low** → Image may be over-optimized or placeholder-like
- **If large file size detected** → Run **webperf-media:Image-Element-Audit.js** for format/sizing analysis

### After LCP-Video-Candidate.js

- **If video is LCP** → Run:
1. **webperf-media:Video-Element-Audit.js** (check poster, preload, formats)
2. **webperf-loading:Priority-Hints-Audit.js** (check fetchpriority on poster)
3. **LCP-Sub-Parts.js** (analyze video loading phases)
- **If poster image is LCP** → Treat as image LCP (run image workflows)

### After CLS.js

- **If CLS > 0.1** → Run **webperf-interaction:Layout-Shift-Loading-and-Interaction.js** to separate causes
- **If CLS > 0.25 (poor)** → Run comprehensive shift investigation:
1. **webperf-loading:Find-Above-The-Fold-Lazy-Loaded-Images.js** (images without dimensions)
2. **webperf-loading:Fonts-Preloaded-Loaded-and-used-above-the-fold.js** (font loading strategy)
3. **webperf-loading:Critical-CSS-Detection.js** (late-loading styles)
4. **webperf-media:Image-Element-Audit.js** (missing width/height)
- **If CLS = 0** → Confirm with multiple page loads (might be timing-dependent)

### After INP.js

- **If INP > 200ms** → Run **webperf-interaction:Interactions.js** to identify slow interactions
- **If INP > 500ms (poor)** → Run full INP debugging workflow:
1. **webperf-interaction:Interactions.js** (list all interactions)
2. **webperf-interaction:Input-Latency-Breakdown.js** (phase breakdown)
3. **webperf-interaction:Long-Animation-Frames.js** (blocking frames)
4. **webperf-interaction:Long-Animation-Frames-Script-Attribution.js** (culprit scripts)
- **If specific interaction type is slow (e.g., keyboard)** → Focus analysis on that interaction type

### Cross-Skill Triggers

These triggers recommend using snippets from other skills:

#### From LCP to Loading Skill

- **If LCP > 2.5s and TTFB phase is dominant** → Use **webperf-loading** skill:
- TTFB.js, TTFB-Sub-Parts.js, Service-Worker-Analysis.js

- **If LCP image is lazy-loaded** → Use **webperf-loading** skill:
- Find-Above-The-Fold-Lazy-Loaded-Images.js

- **If LCP has no fetchpriority** → Use **webperf-loading** skill:
- Priority-Hints-Audit.js

#### From CLS to Loading Skill

- **If CLS caused by fonts** → Use **webperf-loading** skill:
- Fonts-Preloaded-Loaded-and-used-above-the-fold.js
- Resource-Hints-Validation.js (for font preload)

- **If CLS caused by images** → Use **webperf-media** skill:
- Image-Element-Audit.js (check for width/height attributes)

#### From INP to Interaction Skill

- **If INP > 200ms** → Use **webperf-interaction** skill for full debugging:
- Interactions.js, Input-Latency-Breakdown.js
- Long-Animation-Frames.js, Long-Animation-Frames-Script-Attribution.js
- LongTask.js (if pre-interaction blocking suspected)

#### From LCP/INP to Interaction Skill

- **If render delay or interaction delay is high** → Use **webperf-interaction** skill:
- Long-Animation-Frames.js (main thread blocking)

> **Note on cross-skill references:** This skill runs in an isolated subagent (`context: fork`). When a decision tree recommends scripts from another skill (e.g., `webperf-loading`, `webperf-interaction`, `webperf-media`), report the recommendation to the user as a next step — do not attempt to execute those scripts directly. The user or the main agent can activate the appropriate skill to continue the investigation.

## Error Recovery

When a script returns `status: "error"`:

- **LCP/CLS/LCP-Sub-Parts/LCP-Trail** → The page may not have finished loading. Ask the user to wait for full load or reload, then re-run the script.
- **INP** (`getINP()` returns error) → No interactions have been recorded yet. Remind the user to interact with the page, then call `getINP()` again.
- **LCP-Image-Entropy** → No images with measurable BPP found. This is normal for text-only pages or pages where all images are data URIs.
- **LCP-Video-Candidate** → No LCP entries found; see LCP error recovery above.

## Visual Highlighting

By default, scripts highlight the LCP element(s) with colored dashed outlines — useful when the user is watching the browser while the agent runs. To disable:

```js
window.__cwvHighlight = false;
// then run any LCP script
```

Scripts that support this flag: `LCP.js`, `LCP-Sub-Parts.js`, `LCP-Trail.js`.
Loading