Skip to content

Adonis0123/adonis-skills / skills/web-performance-audit

web-performance-audit

Verified

Run a read-only, evidence-led runtime performance audit of a real web application, especially authenticated editors, dashboards, media tools, long-lived SPAs, and pages with many tabs or states. Use when the user asks to profile slowness, inspect every tab, use Chrome DevTools or Chrome Recorder, run React Scan, compare normal and throttled interactions, find memory or rendering risks, or produce a performance report and recommendations without changing code. Discover available browser capabilities, guide the user through official installation when an important capability is missing, route to validated fallbacks, isolate measurement pollution, and distinguish measured facts from screen observations, hypotheses, and unverified paths. Do not use for a Lighthouse-only SEO/accessibility audit, tool installation with no audit goal, or when the user already wants a specific optimization implemented.

showcase.yaml

View on GitHub
hero:
  title: Performance audit in action
  badge: Synthetic, evidence-safe demo
  summary: >-
    This is not a Lighthouse screenshot. The agent freezes 16 interaction surfaces,
    replays the same journey while changing one condition, separates product cost
    from instrumentation noise, and turns the evidence into a retestable backlog.
    Every number below is synthetic and exists only to demonstrate the workflow.
  keyMoments:
    - Reconcile 16/16 primary and nested surfaces instead of treating one screenshot as coverage
    - Compare the same click at 82ms, 184ms, and 296ms to isolate playback and constrained-device risk
    - Discard instrumented timing when React Scan changes the result, while retaining localization value
    - Attach an equivalent retest, success criterion, and stop condition to every recommendation

sections:
  - id: lock-the-scene
    title: Scope · freeze all 16 interaction surfaces
    intro: >-
      Lock the exact target, representative data, initial state, and coverage fingerprint
      before collecting numbers so another engineer can replay the audit.
    messages:
      - role: user
        content: |-
          This media workspace is large. Audit Media, Text, Elements, Library, and every
          nested tab with Chrome DevTools, Recorder, Computer Use, and React Scan.
          Keep the audit read-only and do not change code.
      - role: assistant
        content: |-
          **Audit contract**

          | Field | Frozen value |
          | --- | --- |
          | Dataset | 47:30 synthetic media workspace |
          | Initial state | Synced · Paused · Media / Video · Timeline collapsed |
          | Coverage | 4 primary tabs + 12 nested states |
          | Boundary | Read-only · no save · no generation · no configuration changes |
        toolCalls:
          - name: Browser
            summary: Reconcile visual navigation, the DOM/accessibility tree, and the page manifest
            result: 16/16 surfaces matched; source order and state fingerprint frozen
          - name: Chrome DevTools
            summary: Capture an idle baseline without Recorder or React Scan
            result: Clean baseline captured; no temporary probe left behind

  - id: controlled-replay
    title: Replay · change one condition at a time
    intro: >-
      Recorder fixes the journey while DevTools measures it. Paused, playing, and CPU-throttled
      runs stay separate so correlation is not promoted to causation.
    messages:
      - role: assistant
        content: |-
          **Synthetic median for the same Library → Media click**

          | Comparison | Click duration | Long tasks | Interpretation |
          | --- | ---: | ---: | --- |
          | Paused · normal CPU | 82ms | 0 | clean baseline |
          | Playing · normal CPU | 184ms | 3 | playback contention |
          | Playing · 4× CPU slowdown | 296ms | 5 | constrained-device risk |
        toolCalls:
          - name: Chrome Recorder
            summary: Record and replay the same 12-second tab journey
            result: Replay passed 3/3 times with the same final-state assertion
          - name: Chrome DevTools
            summary: Capture Event Timing, tasks, Network, and recovered heap signals
            result: Synthetic playback window was 52% main-thread busy; the paused control recovered

  - id: isolate-attribution
    title: Attribution · separate product cost from measurement noise
    intro: >-
      Tool output is not automatically product evidence. Start with a clean trace, then localize
      React work, media resources, and third-party scripts without merging incompatible signals.
    messages:
      - role: assistant
        content: |-
          **Three signals, three different claims**

          - `VERIFIED`: synthetic playback occupied about half of the main thread in a repeatable 12-second window.
          - `INFERENCE`: 19 offscreen source images imply a 64.8MiB nominal RGBA8 proxy, not decoded or GPU residency.
          - `MEASUREMENT_POLLUTION`: React Scan changed the same action by 31%, so its timing is discarded.
        toolCalls:
          - name: React Scan
            summary: Replay only the slowest interaction to localize repeated React work
            result: Three high-frequency commit subtrees found; instrumented timing excluded
          - name: Source inspection
            summary: Verify ownership of playback ticks, panel lifecycle, and subscriptions
            result: Three source locations aligned with runtime signals; all other causes remain unverified

  - id: bootstrap-capabilities
    title: Missing tools · use a fallback before asking to install
    intro: >-
      The skill does not assume an MCP server, extension, debugging port, or browser profile.
      Configuration is not considered ready until discovery, handshake, and a harmless call all pass.
    messages:
      - role: user
        content: Chrome DevTools MCP and Computer Use are unavailable. Do not install anything yet.
      - role: assistant
        content: |-
          **Continue now:** use the structured browser for reachable controls and a supported
          PerformanceObserver probe for Event Timing, Long Tasks, and resources. Mark trace call trees,
          Memory, Recorder UI, and visual-only controls as `UNVERIFIED`.

          **Install only if the gap blocks the goal:** show the official source, affected client
          configuration, browser-data exposure, download/restart impact, and the three-part acceptance test.
        toolCalls:
          - name: Capability discovery
            summary: Discover browser, DevTools, visual-control, and script capabilities on this host
            result: 2 available · 2 unavailable · 0 silent installations

  - id: decision-ready-report
    title: Delivery · produce a backlog that can prove its own value
    intro: >-
      The output should answer what to improve first, how to prove the improvement, and when to stop.
      Running many tools is not itself a deliverable.
    messages:
      - role: assistant
        content: |-
          **Decision-ready backlog**

          | Priority | Evidence signal | Smallest direction | Retest / stop condition |
          | --- | --- | --- | --- |
          | P0 | Playback keeps the main thread busy | Separate renderer, timeline tick, React, and decode cost | Stop if production does not reproduce |
          | P1 | Tab switch worsens during playback | Pause inactive subscriptions without losing state | Stop if state breaks or gain is noise-sized |
          | P1 | Offscreen source pixels are excessive | Verify currentSrc and residency before variants/windowing | Stop if resources are not resident |
          | P2 | 79 third-party requests | Compare production tracking on/off | Do not claim UX gain if only count changes |

          **Restoration:** Recorder, React Scan, and CPU throttling are off. The page is back to
          Synced, Paused, and the initial Media state. Nothing was saved, generated, or edited.
        toolCalls:
          - name: Python ledger
            summary: Validate coverage, evidence semantics, retests, stop conditions, and restoration
            result: Ledger valid · 16/16 covered · 0 fabricated metrics · state restored
rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first.Stars: 194kUpdated Claude CodeCodexVerifiedInstall