The problem
The bug bash is on the calendar for the day before launch. Eight people
join a call, click around for an hour, and file what they notice. Half the
findings are the same three obvious bugs. Nobody tests the long company
name or the empty state, because nobody enjoys it.
Then a customer finds it in the first week.
On Claire Vo’s How I AI
in late October, Figma PM Ezra Mechaber described handing the first pass to an
agent with a browser. He writes a test plan, gives the agent a test
account, and reads its findings before the team meets. He said it caught a
1,000-emoji input failure and a grammar bug in a team name.
What you’ll build
A test plan file and a browser connection for Claude Code. You tell the
agent to work through the plan on your staging site. It opens a real
browser, performs each scenario, tries edge-case inputs, and writes a
findings file with screenshots.
Your team’s bug bash then starts from the findings.
Example output
Bug bash findings, invoice sharing, staging, 10 November
Scenarios run: 18. Passed: 13. Findings: 5. Could not run: 0.
F1 HIGH Share link shows the wrong month to the recipient
Scenario 4. Steps: open October report, Share, copy link,
open link in a signed-out window. Expected October.
Saw September. Screenshot: findings/f1.png
F2 MED Team name with an apostrophe breaks the email subject
Scenario 9, edge input "O'Brien & Sons". Subject line reads
"O'Brien & Sons shared a report".
Screenshot: findings/f2.png
F3 LOW Empty state has no button
Scenario 12. A new account with no reports sees "No reports
yet" and nothing to click. Screenshot: findings/f3.png
Not checked: email delivery (cannot read the inbox), payments.
Prerequisites
- A staging or test version of your product at a web address, with fake
data.
- A test account for it. Not your own account and not a customer’s.
- A mogkit workspace with Claude Code. Setup 101 covers it.
Build it
-
Give Claude Code a browser. In the terminal, inside your
workspace:
claude mcp add playwright npx @playwright/mcp@latest
Playwright is a free tool that lets software control a web browser.
-
Write the test plan. Create launches/invoice-sharing/test-plan.md.
Write scenarios as goals, with the result you expect.
# Test plan: invoice sharing
Site: https://staging.example.com
Account: use TEST_EMAIL and TEST_PASSWORD from my message
## Scenarios
1. Sign in. Expect the dashboard.
2. Open the October report and share it by link. Expect a link.
3. Open that link signed out. Expect the October report, read-only.
4. Revoke the link, then open it again. Expect an "expired" page.
## Edge inputs to try in every text field
- empty
- 1,000 characters
- emoji, and a name with an apostrophe and an ampersand
- a script tag such as <script>alert(1)</script>
## Do not
- change account settings, billing, or anything outside sharing
- send email to any address except the test address
-
Start Claude Code and give it the job. Type claude, then:
Work through launches/invoice-sharing/test-plan.md in the browser.
Test account: qa-bot@example.com / the password I will type next.
For every scenario, do it, compare with the expected result, and
try each edge input where a text field exists.
Write launches/invoice-sharing/findings.md with: a summary line,
then one entry per finding with severity, scenario number, exact
steps, expected, actual, and a screenshot saved in findings/.
List anything you could not check. Follow the "Do not" list.
Report only. Do not fix anything.
-
Watch the first few scenarios. A browser window opens and you can
see every click. Stop it if it wanders outside the plan.
-
Reproduce the HIGH findings yourself. Follow the steps in the
file. An agent can misread a page, so confirm before you file a bug.
-
Send the findings file to the team before the bug bash. Ask them
to spend the hour on what the agent cannot judge: does the flow make
sense, is the wording clear, does it feel finished.
-
Run it again after the fixes. Same command, same plan. Compare the
two findings files.
How it works
- The browser connection. Playwright MCP gives Claude a set of
actions: open a page, click, type, read what is on screen, take a
screenshot. Claude uses the same pages your users do. It is not
reading your code.
- The plan is the instruction. The agent is only as thorough as the
plan. Scenarios written as “expect X” give it something to compare
against. A plan that says “test sharing” produces a vague report.
- Edge inputs are where it earns its keep. A person tests the long
name once. The agent tests it in every field of every scenario without
getting bored.
- The “Do not” list is the guard rail. An agent with a browser and a
login can do anything that account can do. The list, the test account
and the staging site together limit the damage a mistake can cause.
- Screenshots make it checkable. Each finding has evidence you can
look at in five seconds, which is how you catch the agent’s own
mistakes.
Variations & next
- Run
launch-readiness first. It lists the gaps in the launch as a
whole: owners, rollback, support briefing. Add its findings to the test
plan as scenarios.
- Size the launch with
launch-tier. A small release may need only
the agent pass. A major one needs the agent pass and the full team.
- Use a different agent. Mechaber’s setup used OpenAI’s Codex with
its browser. Claude in Chrome also works. The test plan file stays the
same.
Limits & honesty
The agent checks behavior against what you wrote. It will not tell you a
flow is confusing, a label is misleading, or the feature solves the wrong
problem. People do that.
It makes mistakes. It can report a slow page as a failure or miss a
visual bug that is obvious to a person. Reproduce anything you plan to
act on.
It cannot check what it cannot see: email delivery, SMS, payments that
need a real card, or anything behind a CAPTCHA.
Use staging and a test account every time. Do not give an agent a login
that can reach real customer data, and do not type the password for a
real account into a chat.