Enforces design review gate after brainstorming — bridges superpowers:brainstorming into the metaswarm quality pipeline
日本語の概要は準備中です。原文の説明を表示しています。
Take screenshots of web pages and UI using Playwright for visual review and iteration
インストールする前に、エージェントに与えられる指示の中身を確認できます。
Agents cannot see rendered web pages, presentations, or UI. This skill bridges that gap by using Playwright to capture screenshots of web pages (local files, localhost servers, or deployed URLs) so agents can visually inspect their work, identify layout issues, and iterate.
This is especially useful for:
This skill requires Playwright to be available. Before first use, check and install if needed:
# Check if Playwright is available
npx playwright --version 2>/dev/null
# If not available, install it (one-time setup)
npx playwright install chromium
Important: Only the chromium browser is needed. Do not install all browsers — it wastes disk space and time.
If npx playwright fails entirely, the user may need to install it:
npm install -g playwright
npx playwright install chromium
npx playwright --versionnpx playwright install chromium and inform the usermkdir -p /tmp/visual-reviewnpx playwright screenshot \
--browser chromium \
--viewport-size "1456,816" \
--wait-for-timeout 3000 \
"<URL_OR_FILE_PATH>" \
/tmp/visual-review/page.png
Parameters:
--viewport-size "1456,816" — 16:9 aspect ratio, good for presentations--wait-for-timeout 3000 — wait 3 seconds for fonts, images, and animations to settlefile:///absolute/path/to/file.html for local filesCapture each slide by appending the slide hash to the URL:
# Capture all slides (adjust count as needed)
for i in $(seq 0 15); do
npx playwright screenshot \
--browser chromium \
--viewport-size "1456,816" \
--wait-for-timeout 2000 \
"<BASE_URL>#/$i" \
"/tmp/visual-review/slide-$i.png"
done
Tip: To capture slides with all fragments visible, add ?fragments=true or use JavaScript injection via a Playwright script. The simpler approach is to accept that fragment-heavy slides will appear in their initial (pre-click) state.
# Mobile
npx playwright screenshot --viewport-size "390,844" "<URL>" /tmp/visual-review/mobile.png
# Tablet
npx playwright screenshot --viewport-size "768,1024" "<URL>" /tmp/visual-review/tablet.png
# Desktop wide
npx playwright screenshot --viewport-size "1920,1080" "<URL>" /tmp/visual-review/desktop.png
Use the Read tool to analyze each screenshot:
Read /tmp/visual-review/slide-0.png
Read /tmp/visual-review/slide-1.png
...
Showing screenshots to the user: The agent can see screenshots via Read, but the user cannot. When the user wants to see what you see — or when you want to discuss a visual issue together — there are two approaches depending on the environment:
Local machine (has a display):
# Open a specific screenshot
open /tmp/visual-review/slide-2.png
# Open all screenshots at once
open /tmp/visual-review/slide-*.png
Remote/headless machine (SSH, tmux, etc.):
Serve the screenshots directory over HTTP so the user can view them in their local browser:
# Start a simple file server (runs in background)
# Use a high port (>10000) to avoid conflicts with dev tools
python3 -m http.server 18080 --directory /tmp/visual-review &
# User can now browse to:
# http://<hostname>:18080/
# http://<hostname>:18080/slide-2.png
# etc.
Pick any open high-numbered port (18080, 19090, etc.) — low ports like 8080 are often taken by dev tools. This serves the entire /tmp/visual-review/ directory with an auto-generated file listing. The user can click through all screenshots in their browser. Stop the server when done:
# Stop the background server
kill %1
# Or find and kill by port
lsof -ti:18080 | xargs kill
Use whichever approach fits the environment. This is useful when collaborating on visual issues, when the user asks what something looks like, or when you want confirmation on a subjective design choice.
For each screenshot, evaluate:
Create a findings summary:
## Visual Review — [Project Name]
### Reviewed: [date]
| Slide/Page | Status | Notes |
|-----------|--------|-------|
| slide-0 | Good | Title centered, accent color correct |
| slide-1 | Issue | Text overflows right edge on mobile |
| ... | ... | ... |
### Issues to Fix
1. [Description of issue + which slide/page]
2. [...]
### Looks Good
- [Things that are working well]
After making fixes:
#/0 for the first slide, #/1 for the second, etc.#/horizontal/vertical format (e.g., #/2/1)class="fragment" is hidden until clicked. Screenshots show the initial state. This is expected behavior, not a bug.S key)--wait-for-timeout 3000 to ensure CDN-loaded themes and fonts are ready| Mistake | Why It's Wrong | Do This Instead |
|---|---|---|
| Not waiting for page load | Fonts/images may not render | Use --wait-for-timeout 3000 |
Using http:// for local files | File URLs need file:/// protocol | Use file:///absolute/path |
| Installing all Playwright browsers | Wastes time and disk | Only install chromium |
| Treating fragment-hidden content as a bug | Fragments are designed to reveal on click | Note it as expected |
| Forgetting to create output directory | Screenshots fail silently | Always mkdir -p /tmp/visual-review first |
| Taking too many screenshots at once | Floods context window | Review in batches of 4-6 |
Not offering to open when discussing visuals | User can't see what you're describing | Use open when collaborating on visual issues |
Upstream: Any skill that produces visual output (presentations, web UIs, email templates)
Downstream: Fixes feed back into the source files, then re-capture to verify
まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Enforces design review gate after brainstorming — bridges superpowers:brainstorming into the metaswarm quality pipeline
日本語の概要は準備中です。原文の説明を表示しています。
Create comprehensive GitHub issues with TDD plans, acceptance criteria, and agent instructions for autonomous PR lifecycle management
日本語の概要は準備中です。原文の説明を表示しています。
Automatic review gate that runs after brainstorming completes - spawns PM, Architect, Designer, Security, and CTO agents in parallel, iterates until all approve
日本語の概要は準備中です。原文の説明を表示しています。
Delegate implementation and review tasks to external AI CLI tools (Codex, Gemini) with cross-model adversarial review
日本語の概要は準備中です。原文の説明を表示しています。
Address PR review feedback systematically — fetch inline comments, review bodies, handle outside-diff-range comments, resolve threads with proper attribution
日本語の概要は準備中です。原文の説明を表示しています。
Analyze the current session and write a self-contained handoff document so a fresh agent can resume the work with full context — outputs a single "Read XXX.md and do YYY." sentence
日本語の概要は準備中です。原文の説明を表示しています。