qa-skill
0xherve/qa-skill
A skill that lets your agent drive a browser to explore, draft, and run test cases, with a human in the loop.
Instalação
npx convez-ai 0xherve/qa-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
QA Skill
A skill that turns an AI agent into a manual-QA assistant. The agent drives a browser to explore a web app, draft test cases, and run them — doing the repetitive, mechanical parts of manual testing while a human stays in the loop for judgment.
It works with Claude Code and any agent that can load a skill and reach a browser through an MCP server.
What it does
The agent does the menial QA work — navigating the app, writing test cases, executing them, recording results — so a QA engineer can focus on deciding what's a real bug. It never decides that for you: it gives you clear evidence, an honest read, and routes doubt to "Uncertain" for you to settle.
Three things it can do:
| Mode | What it does |
|---|---|
| explore | Navigates the app (whole app or a module), discovers features, and drafts test cases. Drafts only — it doesn't run them. |
| run | Executes existing test cases, records pass/fail/uncertain/skipped, and updates their status. Flags any UI drift or uncovered features in the report. |
| update | Re-explores the app, fixes test cases whose steps drifted, adds cases for new features, retires removed ones. |
It finishes the whole batch and reports at the end — no babysitting each step.
Install
Install into the folder where you do QA, using the skills CLI:
npx skills add 0xherve/qa-skill
This drops the skill into your agent's skills folder (.claude/skills/qa/ for Claude
Code, .agents/skills/qa/ for OpenCode) and pins it in a lock file. Re-run the same
command to update.
Setting up the browser
Claude Code — start with Chrome connected:
claude --chrome
OpenCode or any other agent — using the Browseruse MCP
- Install the Browser MCP plugin, from Browser MCP
- Setup the opencode configuration to make the MCP available for your agent:
npx opencode-browser init
Logging in
The agent never handles credentials. You log in first in your own browser session; the agent takes over from there. If the app isn't logged in, it stops and asks you to log in rather than guessing.
How to use it
/qa [explore|run|update] [scope]
- mode —
explore,run, orupdate. Omit it and the agent asks. - scope —
all, or one or more module names (comma-separated). Omit it and the agent asks.
Examples:
/qa explore all # discover the whole app and draft test cases
/qa run documents # run the Documents test suite
/qa update organisation,reviews # bring two modules' cases back in line with the app
If you just type /qa, the agent asks which mode and scope before doing anything.
How work is organised
You keep one folder per app under test. The agent reads and writes inside it:
test-cases/ # the test suite — markdown, the source of truth
<module>/
README.md # module description + summary table with current status
TC01_<Action>.md # one test case: steps, input, expected result, status
runs/ # run history — JSON, easy to query and diff
YYYY-MM-DD_HHMM/
report.json # run summary (totals, failures, changes, notes)
<module>.json # per-test-case results for that run
The markdown suite holds the current state; the runs/ JSON holds the history.
Each test case is numbered per module (every module starts at TC01), so adding or
removing a module never forces a renumber.
After a run, the agent prints a results table in the chat and saves the run JSON.
Statuses
| Status | Meaning |
|---|---|
| Passed | Verified working, clear evidence. |
| Failed | Confirmed broken — a genuine defect, not intended behaviour. |
| Uncertain | Ran it, but the result is ambiguous; needs a human call. |
| Skipped | Couldn't attempt it (upload, email, OTP) or too little signal to claim anything. |
The agent commits to Pass/Fail only when it's confident; anything shaky becomes Uncertain, anything unobservable becomes Skipped. This keeps the "valid bug" rate high.
Layout
skills/
qa/
SKILL.md # the skill the agent loads
references/
qa-workflow.md # the three modes in detail
test-cases-format.md # markdown structure + writing rules
runs-format.md # JSON run history + report table
run-mechanics.md # skipping, drift, failures, test data
browser-actions.md # MCP-agnostic browser capabilities
excel-format.md # column schemas + Excel reporting/navigation