- Home
- Skills
- Testing & Debugging
- Make Your AI Agent Prove UI Changes With Screenshots
More screenshots
Works with the AI Agents you already use
Make Your AI Agent Prove UI Changes With Screenshots
Needs Python, Playwright, Pillow and Playwright's headless Chromium. This skill serves a page locally, drives the states that matter, saves screenshots and an animation strip, compares a second run, and stops its server.
$5.99
Make Your AI Agent Prove UI Changes With Screenshots
Example session with this skill installed
Example request
I added a search box and a "Open summary" button to the Tidewell page (fictional, in
sample/site/). Typing should list matching accounts; the button should slide a summary panel in. Prove both work on a phone-sized screen.
Command run from this folder (Python 3.12, Playwright 1.60.0, Windows 11):
python scripts/shoot.py sample/site --steps sample/steps.json --out sample/output --viewport 390x844 --strip 6 --step-ms 120 --check-twice
- Read your context and instructions
- Compiled the make your ai
- Generated the UI component
Example output (real run)
UI proof: search box and summary panel (8 Nov 2026)
Command: python scripts/shoot.py sample/site --steps sample/steps.json --out sample/output --viewport 390x844 --strip 6 --step-ms 120 --check-twice
Script output
00-load.png 5970 bytes d2e2463cea17
01-typing.png 11149 bytes febf0a1c88aa
02-panel-open.png 12071 bytes e983fae4d378
strip.png 35883 bytes c287385629cd
deterministic: True
server stopped, port 49736 free: True
(The port number differs on every run; the file hashes are the ones from this run and repeated in the second run.)
| File | State | What it shows |
|---|---|---|
00-load.png | Page loaded | Label "Search accounts", an empty search field, the green "Open summary" button, no panel and no list |
01-typing.png | "tide" typed | Field holds "tide"; three white result cards: Tidewell Savings, Tidewell Current, Tidewell Joint |
02-panel-open.png | Button pressed | The three results stay; a green panel reading "Summary: 3 accounts match." sits below the button at full opacity |
Strip: strip.png, 6 frames at 0, 120, 240, 360, 480 and 600 ms. Frame 1 has no panel (opacity 0); frames 2 to 6 show the panel getting darker and moving up about 24 px in total; frame 6 matches 02-panel-open.png.
Determinism: True (all four images were byte-identical in a second run).
Server: stopped, port 49736 free.
Verdict
works. One thing noticed while building this sample: the first version of 02-panel-open.png caught the panel half faded; the script now waits for finite CSS animations to end before each screenshot, and the Gotchas say so.
00-load.png
TSX · React component
Example file from a real run - the skill writes it into your workspace.
Connects securely to your tools. The creator never sees your data.
About this skill
Needs Python, Playwright, Pillow and Playwright's headless Chromium. This skill serves a page locally, drives the states that matter, saves screenshots and an animation strip, compares a second run, and stops its server.
The proof
This is real output copied from example-output.md:
00-load.png 5970 bytes d2e2463cea17
01-typing.png 11149 bytes febf0a1c88aa
02-panel-open.png 12071 bytes e983fae4d378
strip.png 35883 bytes c287385629cd
deterministic: True
server stopped, port 49736 free: True
The same file records what each screenshot shows and describes all six frames in strip.png.
What you get
SKILL.md: the evidence-first UI workflow for Claude Code and OpenAI Codex.scripts/shoot.py: local server, state driver, screenshots, frame strip and repeatability check.sample/siteandsample/steps.json: a small interactive worked page.sample/output: three state screenshots, a six-frame strip andmanifest.json.tests/test_shoot.py: end-to-end pytest coverage of output files, sizes, repeatability and cleanup.example-task.mdandexample-output.md: the command and visual report.
How to use
- List the states worth proving in a small JSON file.
- Run
scripts/shoot.pywith the site folder, viewport and output folder. - Add
--check-twiceto compare image hashes across two runs. - Open every PNG and write one sentence about what it actually shows.
- Fix any wrong state and rerun until the evidence matches the change.
Known limitations
- Repeatable screenshots prove repeatability, not correctness. A person still has to inspect every image.
- Live data, remote fonts and other external inputs can make images differ.
- Font rendering can differ between operating systems.
- The script writes screenshots and a manifest to the output folder you choose.
FAQ
Does it use my everyday browser? No. It launches Playwright's own headless Chromium.
Does it leave a server running? The script shuts down its own local server and prints whether the port is free.
Can it capture interactions? Yes. A step can click, fill a field, run JavaScript and advance the fake clock.
Which agents does it work with? It is a SKILL.md skill for Claude Code and OpenAI Codex.
Does a clean second run prove the UI is right? No. You must still look at and describe every saved image.
Requires Playwright and Chromium. Font rendering differs by OS. Infinite loops break determinism. Real-time timers require manual advancement via wait_ms.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 4 days ago
Requires