Run an autonomous QA review of a web UI by invoking the local `qa-cli` command via Bash. Trigger when the user asks to QA a URL, check a deployed site for accessibility, usability, or correctness issues, get a structured review of a page, or run a quick visual sanity check on a live page (e.g. "what color is the CTA?").
Installation
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Run an autonomous QA review of a web UI by invoking the local `qa-cli` command via Bash. Trigger when the user asks to QA a URL, check a deployed site for accessibility, usability, or correctness issues, get a structured review of a page, or run a quick visual sanity check on a live page (e.g. "what color is the CTA?").
When the user asks you to QA a web UI / review a deployed page / check a URL for issues:
1. Pick the right subcommand
qa-cli review — full agent loop (~30–90s). Returns a structured ReviewResult JSON with verdict, summary, findings (each with severity, area, issue, suggestion), and steps_taken. Use this for any real review.
qa-cli visual — one-shot visual question, returns a short free-text answer. Use for quick "what color is X" / "does the layout look broken" sanity checks.
2. Invoke
uv run qa-cli review --url <url> --instruction "<what to focus on>"
uv run qa-cli visual --url <url> --question "<single question>"
The CLI logs progress to stderr; the result (JSON for review, text for visual) goes to stdout. Pipe to jq if you want to filter.
3. Surface the result
For review:
Lead with the verdict and the headline summary.
List findings ordered as returned (already high-severity-first). Show each as [severity] area — issue and include the suggestion when actionable.
If verdict == "fail", make it clear what blocked completion.
For visual: just relay the answer.
Notes
Prerequisite: the qa-cli command comes from the computer-use-agents-demos repo. If uv run qa-cli ... errors with "command not found", that repo isn't installed in the current working directory — tell the user to clone it and run uv sync before retrying, or fall back to whichever QA tool is already on PATH.
Requires HAI_API_KEY in .env. On a missing key the CLI exits with error: HAI_API_KEY is not set... — relay that to the user verbatim; don't retry.
review is expensive (one cloud session per call). Don't run multiple reviews in parallel for the same task — call it once with a focused --instruction.
Examples
User: "Please QA https://news.ycombinator.com for accessibility."
You: Run uv run qa-cli review --url https://news.ycombinator.com --instruction "focus on accessibility — alt text, form labels, keyboard nav, contrast". Parse JSON. Surface findings.
User: "What color is the main CTA on https://example.com?"
You: Run uv run qa-cli visual --url https://example.com --question "what color is the main call-to-action button?". Return the answer.
