Three copy-paste prompts that force AI to prove its work. Because the report reads the same whether the work happened or not.
3copy-paste prompts
1installs permanently
2orders it claimed shipped
0alerts that fired
The shortcut: paste the Receipt Check at the end of any AI work session. It forces the AI to re-pull the live state of everything it claimed to do, and tag each claim VERIFIED, SUBMITTED, or FAILED, before you trust a word of it.
Selling From the Beach
The Prompts
page 02
paste these, word for word
Three prompts that force a recheck.
The AI's report is a claim. These prompts make it go back to the live system and bring evidence. One for after any task. One to install permanently. One for a second opinion.
Prompt 1 · use right after the AI says it finished something
This is the one from the essay. Paste it at the end of any work session where the AI touched something real: orders, listings, ads, emails, files.
the receipt checkStop. Before we move on, audit the work you just told me was done.
Go back to the source system and pull the current live state of every action you claimed to complete. Fresh. Not from memory, not from earlier in this conversation.
Then give me a table with one row per action: what you claimed, what the live system shows right now, and a verdict of VERIFIED, SUBMITTED, or FAILED.
Rules:
1. An accepted request is not completed work. If all you have is an acknowledgment, the verdict is SUBMITTED, not VERIFIED.
2. Re-read every response you received for warnings, partial successes, or error fields. Quote them word for word, even the ones that seem minor.
3. List FAILED and SUBMITTED rows first. Verified wins go last.
4. If you cannot reach the system to check, say exactly that and mark the row CAN'T CHECK. Do not guess. Do not fill the gap.
Prompt 2 · install once, applies to every future task
Put this in custom instructions, project instructions, or your agent's memory file. After this, "done" has a definition.
the standing orderFor any task that changes something outside this chat (orders, listings, ads, emails, files, payments), these rules always apply:
1. Never report a task complete because the request was accepted. Complete means you re-checked the live result in the system afterward.
2. Use exactly three statuses in every report: VERIFIED (you re-checked the outcome at the source), SUBMITTED (request accepted, outcome not confirmed), FAILED. Never call SUBMITTED work "done."
3. Quote any warning or partial-success message you received word for word. Never summarize a warning away.
4. Open every report with what failed or could not be verified. Successes go last.
5. If I ask "did it work?", re-check the source before answering. Never answer from memory.
Prompt 3 · use in a brand new chat, zero history
The AI that did the work will defend the work. So don't ask it. Open a fresh chat, paste the first AI's report where marked, and let a stranger grade it.
fresh eyesYou are auditing work that another AI says it finished. You did not do this work. Assume the report is wrong until you find evidence it is right.
Here is the report:
[PASTE THE OTHER AI'S REPORT HERE]
Check every claim against the live system, one by one. Watch for the four ways these reports go wrong: actions that were accepted but never completed, items silently skipped from a list, warnings that got summarized away, and numbers that do not add up to the totals claimed.
Give me a verdict on each claim, with the evidence next to it. Anything you cannot verify is a finding, not a footnote.
The AI will tell you everything went through. Go look anyway.
Selling From the Beach · companion to "It Always Says Done" · June 2026
Subscribe at sellingfrom.co/newsletter. Weekly, free, 6,000+ Amazon sellers.