This is a new companion worksheet based on my two-business experiment. It isn’t the original Polsia prompt or an audit tool from the video.
A useful test checks the result a person needs. A claim that something was deployed, or a count of files created, answers a different question.
Define done before starting
Choose one result, a time or spending limit, and an early review point. Then fill in this table.
| Claim to check | What I will actually do | Evidence to keep |
|---|---|---|
| The page works | Open the exact live address as a visitor and use the main action. | Address, time, and what happened. |
| The directory contains useful records | Compare a named sample with the original sources. | Sample method, record IDs, sources, and discrepancies. |
| The information persists | Save a test item, reload, and retrieve that same item. | The test input and retrieved output. |
Adapt these to your project. A successful server response doesn’t prove that a page rendered or its form worked. A non-empty database doesn’t prove that its records are accurate.
For a small directory, check every record if feasible. If you sample, record how you chose the sample. Don’t turn a convenient handful of successes into a claim that the entire directory is verified.
Ask for evidence
Before starting, turn this brief into three observable acceptance checks.
Brief:
Time or spending limit:
The first review point:
For each check, name the actual page, file, record, or user action you will
inspect. Tell me whether you have access to inspect it directly.
At review, report PASS, FAIL, or UNVERIFIED for each check. Include the input,
expected result, observed result, and evidence. If you cannot inspect the live
result, say UNVERIFIED. Do not substitute a task report for that inspection.
Keep failed checks visible. Ask before expanding scope or exceeding the limit.Read the evidence yourself. Another AI can help review it, but it can also repeat the same mistake. A fresh chat is another critique, not independent proof that the product works.
Keep three questions separate
- Was something built? Inspect the actual artifact.
- Does it work as intended? Perform the checks.
- Does anyone want it enough to use or pay for it? That needs a separate audience test.
My experiment didn’t reach the third question: I left payments disconnected. Don’t interpret that zero revenue as a completed demand test.
For an additional critique prompt, I’ve also published The Second Opinion. It came from a separate news episode, not either Homebrew Club build. Use it to find claims to investigate; its verdict and explanations still need checking against sources.