Hand your prompt to someone else. If the result changes, it is not finished.

,
By Cheri Stockton, Chief Technical Therapist, Hot Hand Media.

The prompt handoff test means giving your prompt to another person or a fresh chat with no extra explanation, and checking whether the result stays the same. If the output changes, the prompt is a draft, because the missing instructions are still living in your head. A finished prompt holds without you in the room.

  • A prompt is not a system until it produces the same result in someone else’s hands.
  • The prompt handoff test replaces your opinion of the prompt with evidence of what it actually does.
  • If you keep re-explaining the task, the explanation belongs in the prompt, not in your memory.
  • Writing the SOP was never the same job as getting out of the loop.
  • A prompt that only works when you narrate it is a prompt that still owns your time.

What is the prompt handoff test?

The prompt handoff test is a check where you give your prompt to another person or a blank chat with zero verbal context, run it, and confirm the result matches what you expected. If the output drifts, the prompt is unfinished, because the instructions that steer it still live only in your head. It is the cheapest quality gate you have. You do not need a tool. You need one other human, or a second clean window, and the honesty to look at what comes back.

Here is the line that started this for me. A client said, “So I have, yeah, the SOPs, right? But she just, I just have to keep telling her.” The documents existed. The work still came back to her every single day. The SOP was written. The job was not done.

A prompt is not a system until it produces the same result in a hand it has never been in before.

Why does the work still come back to you?

The work comes back because writing something down and removing yourself from it are two different jobs, and most people stop after the first one. A document captures what you remember to say out loud. It does not capture the dozen small corrections you make in the moment without noticing you are making them. That gap is where the task returns to your desk.

There is a name for this. Polanyi’s paradox is the idea that we can know more than we can tell, meaning you can do a task perfectly while being unable to fully explain the rules you are following. That is exactly why the SOP exists and you still keep re-explaining. The research on this gap is clear. The knowledge in your hands rarely makes it fully onto the page the first time.

Delegation experts say the same thing in plainer words. Handing off work without the full context and a clear objective is not delegating at all. It is a pause before the work boomerangs back.

The prompt handoff test, in order

Run these steps on one prompt. Do not read theory first. Just do the moves.

  1. Pick the prompt you reuse most and copy it exactly as written.
  2. Open a brand new chat so no earlier context is helping it along.
  3. Paste the prompt and add nothing. No clarifying note, no “oh and also.”
  4. Run it and save the output as Result A.
  5. Hand the same prompt to another person and have them run it cold.
  6. Save their output as Result B.
  7. Put Result A and Result B side by side.
  8. Mark every spot where they differ in format, tone, length, or decision.
  9. For each difference, find the instruction you gave verbally but never wrote.
  10. Write that instruction into the prompt in plain words.
  11. Delete any line in the prompt that both results already got right without it.
  12. Run the whole test again with the revised prompt.
  13. Repeat until Result A and Result B match without you saying a word.

If you have to keep telling it, you are the document, and the prompt is just a decoration.

What does a passing result look like?

A passing result is two independent runs of the same prompt producing outputs you would both accept and ship, with no edits that only you know how to make. The format holds. The tone holds. The judgment calls land the same way whether you are watching or not. That is the moment the prompt stops being a draft and starts being a system you can actually leave alone.

Two things worth naming here. First, model updates can shift output on an identical prompt, so a prompt that passes today deserves a re-check after a major model change. The prompt engineering guidance from the makers of ChatGPT recommends pinning versions and testing outputs for exactly this reason. Second, the clearest version of this test already lives in the official docs for Claude, which call it a golden rule. Show your prompt to a colleague with minimal context. If they would be confused, the model will be too.

Draft prompt versus finished prompt

Signal Draft prompt Finished prompt
Needs your narration Yes, every time No
Output across handlers Changes Stays the same
Missing instructions In your head On the page
Who owns the task You, still The prompt

What to do with the differences you find

Every difference between two runs is a buried instruction you have been supplying with your voice instead of your text, and your only job is to move it from one place to the other. You are not rewriting the whole prompt. You are hunting the silent corrections and writing them down one at a time.

Common ones I see. You assume a format but never state it. You assume a reading level but never name it. You assume the thing should stop at three options but you never said three. Each of those is a one-line fix. The prompt gets longer and your involvement gets shorter, which is the whole trade you are after.

If you want a sharper handoff standard, borrow the question delegation coaches recommend. The guidance on effective delegation says to ask the person whether your instructions were actually clear. Run that same question against your prompt. If the honest answer is no, you found your next edit.

The phrase “we can know more than we can tell” comes from chemist and philosopher Michael Polanyi, writing in the 1960s. He used riding a bicycle as the example. You balance perfectly and cannot explain the physics you are using to do it. Your most reused prompt is carrying the same kind of silent knowledge.

Almost every time I open a client’s most-used prompt, I find the same thing. The prompt looks finished and the person swears it works. Then they run it in front of me and narrate three corrections before the output is usable, without noticing they did it. That narration is the real prompt. The written version was never carrying the full load. The handoff test exists to make that narration visible so you can finally write it down and stop being the one holding the task together.

FAQs

How do I know if my prompt is finished?

Your prompt is finished when someone else runs it cold and gets a result you would ship without edits. Hand it over with no verbal context. If the output matches what you expected, it holds. If you feel the urge to explain one more thing, that one more thing is the unwritten part of the prompt.

Why does my prompt work for me but not for my assistant?

It works for you because you are quietly filling the gaps with corrections you make by habit. Your assistant does not have your habits, so the gaps show. The fix is to find each silent correction and write it into the prompt as a plain instruction.

What is the prompt handoff test?

It is a check where you give your prompt to a second person or a fresh chat with no extra explanation and see if the result stays the same. Matching results mean the prompt is finished. Changed results mean instructions are still living in your head instead of on the page.

Why does the same prompt give different answers?

Two reasons. The prompt may be missing instructions you supply verbally, so each handler guesses differently. Or the model itself updated, which can shift output on identical text. Pin your prompt, re-test after major model changes, and write down anything you keep correcting by hand.

Is a prompt the same as a system?

No. A prompt is a set of instructions. A system is a prompt that produces the same result no matter whose hands it is in. A prompt only becomes a system after it passes the handoff test and stops requiring you to narrate it.

How often should I re-run the handoff test?

Run it when you first think a prompt is done, and again after any major model update. If you notice yourself adding a verbal correction during normal use, that is a signal to test and revise. A prompt that passed once can quietly drift back into draft.

Pick your single most-used prompt. Open a clean chat, paste it with no notes, and have one other person run the same prompt. Compare the two outputs and write down every instruction you have been saying out loud. That list is your edit queue.

Save this and run the handoff test on your most-used prompt this week.

Add Hot Hand Media as a Preferred Source

… and see more of my posts in your Google results.

Your business should not run on chaos.

Tell me what you are trying to fix and I will tell you whether Hot Hand Media is the right fit. No pitch deck. No salesperson voice.