Ask for evidence rather than a status: make it run the thing and paste the real output, or take a screenshot. "Done" from an agent usually means "written", sometimes means "ran once", and rarely means "checked". Measured, generated code is syntactically correct well over 90 percent of the time and secure only about half — so working and correct are genuinely different questions.
The prompt
Show me evidence this works. Run it and paste the actual output. If there is a UI change, take a screenshot.
| When to use it | Before you believe anything is finished. Every time the agent says it is done. |
|---|---|
| Why it is worded that way | "Evidence" and "actual output" are doing the work. Without them you get a description of what the code should do, which is a restatement of the intention you already had. Pasted output is a fact you can check. The screenshot clause exists because UI changes are where confident descriptions and reality diverge most often. |
Why syntax correctness is the wrong reassurance
Veracode's Spring 2026 testing across more than 150 models and 80 coding tasks found syntax correctness above 95 percent while the security pass rate sat near 55 percent — and two years of model releases moved that second number essentially nowhere. Code that runs is now the easy part. Code that is right is the job, and the job is verification.
The harder follow-up
Once you have output, ask it to attack its own work: "Review this change as though you did not write it. What breaks if the input is empty, enormous, or sent twice? What did the author assume that they should not have?" The empty / enormous / twice trio catches most real-world input failures.
The one that catches drift
On long sessions an agent gradually answers a slightly different question than the one you asked. Before calling anything finished: "Read the original request again, then tell me honestly what it does not yet do that I asked for."
Putting it in the rules
Write it down once: "Done means you ran it and pasted the actual output, or took a screenshot. Not a description." After that you stop having the conversation and start getting the evidence.
The Operator Prompt Pack
This is one of 60 prompts, from the phase "Verify like you did not write it". The pack has the rest, each with the same reasoning underneath, plus the research they are built on and a house-rules file you paste in once.
Tools this applies to
| Tool | What it does | Price |
|---|---|---|
| Claude Code | An AI agent in your terminal that reads your files, edits them, runs commands and deploys. | ~€17–20/mo |
| Cursor | A code editor with AI. The step between browser builder and terminal. | ~$20/mo |
Every tool has its own page with what it costs and who should skip it. Some links are affiliate links.
More prompts like this
The prompt that stops an AI agent going in circles
What do you do when an AI coding agent keeps trying fixes that do not work?
The prompt to run before any app with a database goes live
How do you check whether your database is exposed before launching an app?