A prompt somebody else can run

The last step, and the one that proves you understood

You have a prompt that works. You give it to a colleague. It works less well for them, and neither of you can say why.

Making a prompt transferable is the final skill in this course, because it forces every implicit thing into the open, and the implicit things are where the quality was.

The four things that do not travel

Your context. The prompt worked because you always paste the same kind of invoice, from the same supplier, in the same format. Your colleague pastes something else.

Your standing brief and memory. Half your quality may be coming from custom instructions your colleague does not have. This is the commonest cause of a prompt that “does not work for me”, and it is invisible from both sides.

Your judgement about the output. You know that a NOT_STATED in the VAT column means go and look at the paper copy. Your colleague sees a completed table.

Your product setup. Your tool has search and file reading; theirs is a different tier or a different product entirely.

What to hand over

Not a prompt. A short document:

## Supplier quote comparison
For: comparing two or more quotes on the same criteria.
Tested: 2026-09-04, [model], works.
Needs: nothing special — plain chat, no search required.

PROMPT
[the prompt, with slots in CAPITALS]

WHAT TO PASTE
Both quotes in full, including the terms page. Not summaries.

WHAT GOOD LOOKS LIKE
One row per supplier. Every figure has a quoted source line.
Missing figures say "not stated" — that is correct, not a failure.

WHEN TO DISTRUST IT
- If any row has no source quote, check that figure by hand.
- If the two quotes are on different tax bases, the comparison is
  invalid; check this first.
- Lead times are invented if the "do not estimate" line is removed.
  Do not remove it.

IF IT GOES WRONG
Check the quotes were pasted in full. Then check whether your
custom instructions are adding anything.

Six headings, ten minutes to write, and it survives you leaving the organisation.

The when to distrust it section is the one nobody writes and the one that matters most. A prompt handed over without it is a machine with no dial markings.

Test it on a real person

Give it to a colleague and watch them use it once, without helping. You will learn more in five minutes than from any amount of re-reading.

What you will see: they paste a summary instead of the document. They stop at the first output rather than reading the source column. They remove the odd-looking line because it looks like clutter. Each of those is a fix to the document, not to them.

The team version

Three files, in a place everybody can reach: the prompts, the examples, and the test items. Plain text, not a system that needs maintaining.

Two rules keep it alive: every prompt has an owner, and every prompt has a date. An unowned prompt library becomes a graveyard of things nobody trusts within about six months, and the failure is silent: people stop using it rather than reporting that it broke.

What this actually teaches you

Writing the handover forces you to say what the prompt assumes, where it fails, and how you would know. Those are exactly the things you were carrying in your head and calling skill.

If you cannot write the “when to distrust it” section, you do not yet know when to distrust it, which is worth finding out while it is still your own work rather than somebody else’s.

The end of the course, in one paragraph

You have three files: prompts that worked, examples from your own archive, and ten test items. You know that the model receives one flat block of text and nothing else, that agreement is its cheapest output, that fluency is not evidence, and that the checking has to be designed before the prompt is sent rather than performed after the answer arrives. Everything else here was an elaboration of those, and the files are what turn them into a practice rather than a set of things you once read.