agent-first-screenshots — Claude Code skill: clean screenshots via CDP
The agent-first-screenshots agent drives the live app over CDP and keeps a frame only after structural, pixel and vision checks pass. Built for landing pages, newsletters and decks.
LicenseRef-Proprietary ★ 7,523 updated 2026-10-29
npx skills add Devin-AXIS/iPolloWork --skill agent-first-screenshots -a claude-codeWhat agent-first-screenshots does
Reload the page, walk the UI the way someone prepping a demo would, hide a stray badge or footer link with display:none, capture. A DOM check runs first: the scroll area has to be taller than 100px, hero copy sits inside the viewport, nothing modal is left open. Fail that and you reload instead of patching styles. Overriding flex on structural containers collapses the layout, so the skill forbids it outright.
Then the PNG is decoded and sampled at the hero rects. Variance is the signal, not background share: filled regions sit above ~200, flat ones below 50, and a whole-image background ratio past 0.97 means an empty frame. Stats still miss overlap and clipping, which is why a vision model grades the image against a JSON rubric; if any_obvious_defect comes back true, the frame is rejected no matter what the first two layers said.
Two scripts come with the skill: screenshot-verify.mjs for an existing PNG, and capture-verify.mjs, which shoots at 2x, checks the regions and saves only passing frames. Both lean on sharp. The vision gate is mandatory, so an agent with no access to a model that looks at images gets half the pipeline. Runs in Claude and Codex. This is not pass/fail evidence either; the skill points to fraimz for that.
What it helps with
- capture clean app shots for a landing page
- regenerate screenshots after a redesign without hand-cropping
- catch clipped text and collapsed layout before publishing
- fit a frame to a target aspect ratio with a CDP metrics override
- reject blank frames before they reach a deck
Example prompts for the agent
- Take fresh screenshots of the app for the pricing page at 1440x900 with
deviceScaleFactor: 2and verify each frame before saving. - Redo the dashboard screenshots after the redesign and drop any frame that fails the vision check.
- Run
capture-verify.mjsover the settings page and keep only the PNGs that pass.
How to install agent-first-screenshots
1Skills CLI (Vercel)
npx skills add Devin-AXIS/iPolloWork --skill agent-first-screenshots -a claude-code2Manually: put the files in place
~/.claude/skills/agent-first-screenshots/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/agent-first-screenshots/
1Skills CLI (Vercel)
npx skills add Devin-AXIS/iPolloWork --skill agent-first-screenshots -a codex2Manually: put the files in place
.agents/skills/agent-first-screenshots/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Devin-AXIS/iPolloWork --skill agent-first-screenshots -a cursor2Manually: put the files in place
.cursor/skills/agent-first-screenshots/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Devin-AXIS/iPolloWork .opencode/skills/agent-first-screenshots2Manually: put the files in place
.github/skills/agent-first-screenshots/SKILL.md1Skills CLI (Vercel)
npx skills add Devin-AXIS/iPolloWork --skill agent-first-screenshots -a gemini-cli2Platform CLI
gemini skills install https://github.com/Devin-AXIS/iPolloWork --path .opencode/skills/agent-first-screenshotsManually: put the files in place
.devin/skills/agent-first-screenshots/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about agent-first-screenshots
How do I install agent-first-screenshots in Claude Code?
Run npx skills add Devin-AXIS/iPolloWork --skill agent-first-screenshots -a claude-code. The skill lands at ~/.claude/skills/agent-first-screenshots/SKILL.md.
What does agent-first-screenshots do?
It drives the real app through CDP and releases a screenshot only after DOM, pixel-variance and vision checks clear it.
Is it a replacement for end-to-end test evidence?
No. It produces product imagery; for pass/fail proof the skill directs you to fraimz.