You can do some pretty cool things with this 4th step. And it goes beyond just a verification step with your coding agent because it’s gonna miss most of its errors anyway.
Something I’ve been doing is creating test suites to give to ChatGPT agent. First, I give the following prompt to Claude or whatever is building the app to get me tests. Here’s the prompt
———————-
{
"role": "Front End User Test Generator",
"task": "Generate a structured testing plan that ONLY covers what a single end user can manually test through the front end UI. Do not include automated tests, backend checks, or code-level validation. Focus entirely on what a real person can see, click, enter, or experience in the browser or app interface.",
"projectFeatures": "{{INSERT_FEATURE_LIST_OR_PROJECT_DESCRIPTION_HERE}}",
"outputFormat": [
{
"testCaseID": "TS-###",
"testType": "UI | Interaction | Flow",
"feature": "{{Feature Name}}",
"scenario": "{{Brief description of what the user is doing in the UI}}",
"steps": ["Step 1", "Step 2", "Step 3", "..."],
"expectedResult": "{{What the user should see or experience if the test passes}}"
}
],
"rules": [
"Each testCaseID must be unique and sequential (e.g., TS-001, TS-002...).",
"All tests must be written from the perspective of a single user interacting with the front end only.",
"No backend, API, or automation steps are allowed.",
"Prioritize real-world user flows, edge cases, and visible outcomes.",
"Use plain, non-technical language so that anyone could follow and execute the tests."
]
}
——————————
Then I copy paste the output I get back to ChatGPT agent along with a url and tell it to get started. Here is the prompt I send along with the test suite and url.
—————-
You are in Ultra-Thorough Implementation Mode. Begin now.
You have already received a structured test suite output. Begin executing the test cases one-by-one with zero omissions or summarization.
For each test case, follow this exact format:
Step [#]: Execute Test Case TS-### – [Short Test Title]
Tools needed: [List tools or utilities used for this specific test]
Preconditions: [All setup states or conditions required before execution]
Configuration Required: [Describe any necessary configuration, environment, or state]
Output: [Describe the result of executing the test]
Success criteria: [Define the condition for a passing result]
Always process sequentially by testCaseID (e.g., TS-001, TS-002, etc.).
Do not skip, merge, or summarize any step.
Explicitly log expected vs. actual result comparisons.
For any deviation or failure, capture exact repro steps and environment state.
Continue until every test case in the suite is executed and validated.
Begin executing the test suite now. Do not stop. Do not simplify. Do not summarize.