| name | visual-verify |
| description | Use when about to declare done on any output with a visual surface -- an HTML page, a rendered UI, a diagram, a slide, a screenshot-able view -- to confirm by looking at the rendered result instead of asserting it. Also use when asked to screenshot or visually check a page or a local file. |
| metadata | {"user-invocable":true,"type":"technique"} |
Visual Verify
Visual verification is the discipline of confirming a visual output by
rendering it and looking at the pixels, not by asserting the markup "should"
look right. When work produces something a person would see -- a page, a UI, a
diagram, a slide -- the honest check is a screenshot the agent then reads with
its own image-reading tool. An agent that ships visual output without looking
at it has not verified; it has hoped.
This skill teaches the disposition and hands you a reference implementation
(scripts/screenshot.cjs) as support, not as a spec. It is generic -- any HTML
or URL, any project, inside Gaia or out. For the response contract and the
verification block this feeds, see agent-protocol.
Core principle
Three judgments shape a good visual check. Reason through each; do not follow a
fixed recipe.
- Look, and look where it breaks. Markup that parses is not layout that
works. Text clips, boxes collide, a column that is fine at 1440px overflows
at 380px, a palette that reads in light mode fails in dark. So render across
a spread of widths (desktop down to narrow mobile) AND across the themes the
output supports (light/dark) -- one viewport in one theme is not a
verification of a responsive, themed surface. Reading the images is the
verification; the script exiting 0 is not.
- Find the browser where it lives; obtain it only if it is truly absent.
Playwright caches its browsers in a location that varies by OS and by the
PLAYWRIGHT_BROWSERS_PATH override -- do not assume one fixed path. Locate
the browser dynamically (resolve the highest chromium-* build present) and
launch it via an explicit executablePath, so you use what is already on
disk. If nothing is present anywhere, obtaining a browser is a legitimate,
deliberate one-time step -- a real install you run with approval. The
friction to avoid is not installing per se; it is letting a tool silently
fetch a mismatched revision when a perfectly usable browser already sits in
the cache.
- Put the captures where their context wants them. The output location is a
decision, not a fixed folder. Read the context and choose:
- Backing a brief (inside Gaia): write them where that brief's evidence
lives, so the capture becomes part of the audit trail.
- Generic / outside a brief: a temporary directory -- it keeps the
workspace clean and is not a meaningful mutation -- or the location the
user or context names.
The reference script takes the out-dir as an argument precisely so this
decision stays with you.
Process
-
Have the output at a URL a browser can open -- a local file as
file:///absolute/path, or a running dev server as its http://localhost:PORT.
-
Decide the output location by the criterion above (brief-evidence vs
temp vs given), then locate the browser where it lives.
-
Capture across the widths and themes that matter for this surface. The
reference script does this:
node <skill-dir>/scripts/screenshot.cjs <url-or-file> <out-dir> [widths] [colorSchemes]
Defaults to 1440,900,700,500,380; pass light,dark as the fourth argument
to capture both color schemes. An app with a bespoke theme toggle
(localStorage/class rather than prefers-color-scheme) may need a
project-specific step -- adapt the reference, do not assume it fits.
-
Read every screenshot with the Read tool and actually look: overflow,
clipping, collisions, contrast in each theme, and whether the change you made
is visible and correct. Skipping this and trusting the exit code is the
failure mode the whole skill exists to prevent.
-
Report what you observed. In the verification block use
method: "self-review" (visual inspection is a genuine self-review), list
the widths/themes checked in checks, and put the PNG paths and concrete
observations in details. Saw a defect -> result is not pass; loop and
fix.
Why the reference script is shaped this way
The explicit-executablePath + dynamic-revision approach exists to use the
browser already present instead of triggering a redundant fetch. Two
alternatives are rejected by evidence -- know why before reaching for them:
- The one-line
playwright screenshot CLI. It targets the revision the
installed package pins, not the one on disk; when they differ it silently
triggers a browser download and stalls the check on an approval prompt. The
reference script sidesteps this by launching what is cached.
chrome-devtools-mcp. It does not support WSL and requires a live Chrome
instance, so it is not a portable verification path here.
Anti-patterns
- Declaring done without looking. Emitting
verification.result: "pass" on
visual work because the script ran is the exact hollow pass this skill
prevents. The pass is the image you read, not the exit code.
- Assuming one browser path or one revision. A hardcoded cache path or
chromium-1223 in a launch line is a time-bomb; the location varies by
environment and the cache is repopulated over time. Resolve both dynamically.
- Fetching a browser reflexively when one already exists. The reflex to run
a fetch on any hiccup converts a free check into an approval wall. Use what is
on disk; obtain deliberately only when nothing is present.
- Checking one width in one theme. Layout and contrast fail at the edges --
narrow viewports and the non-default theme -- not at the comfortable default.