The file that makes this compound

What “prompt engineering” means for most working people

Not a job title. A small file of prompts that produced good output, with the specifics swapped out next time.

Each one took you a few minutes to work out and saves those minutes every week afterwards. Nobody else can write them for you, because they encode what your work actually needs.

What goes in an entry

Not just the prompt. Five things, and the extra four are what make it usable in three months:

## Supplier quote comparison
Last checked: 2026-09-04 (Claude, worked)
Use when: two or more quotes to compare on the same criteria.

PROMPT:
Compare the quotes below on: unit price, minimum order, lead time
in days, delivery to [CITY]. One row per supplier. Quote the line
each figure comes from. Where a quote does not state something,
write "not stated" — do not estimate. Then one line: which is
cheapest for [QUANTITY] units, and what would change the answer.
<quoteA>[PASTE]</quoteA>
<quoteB>[PASTE]</quoteB>

Notes: without the "do not estimate" line it invents lead times.
Tested on: quotes from Adeyemi and Okonkwo, Aug 2026.

The date and model tell you whether to re-test. The use when stops you reading five prompts to find the right one. The notes stop you removing the line that is doing the work: this is the entry people regret leaving out, because a bare prompt looks like it contains removable padding. The slots in capitals make it obvious what to replace.

Slots, not variables

[CITY], [QUANTITY], [PASTE]. Capitals in brackets, because they are visible when you skim and impossible to leave in by accident: an unreplaced [CITY] in your output is obvious, whereas a leftover “Nairobi” from the last time you used it is not.

That is not a small point. The commonest error in reusing a prompt is a stale specific from its previous use, and slots are the only reliable defence.

Where to keep it

A plain text file or a note. Not a system that needs maintaining.

The only real requirement is that you can find it in five seconds from wherever you work. A file you have to open a tool to reach is a file you will stop using, and a prompt library that goes unused is worse than none, because you think you have one.

Once you have more than about fifteen entries, split by area of work rather than alphabetising. You look for prompts by situation, never by name.

What to put in and what to leave out

In: anything you have run three times. Anything that took more than ten minutes to get right. Anything with a non-obvious line in it: a constraint you discovered you needed.

Out: one-off questions, anything you would rewrite from scratch anyway, and any prompt you cannot say what it is for. If the “use when” line is hard to write, the entry is not a tool yet.

Sharing it

A prompt shared without its notes and its test items usually disappoints, which is the subject of a later lesson. If you hand one to a colleague, hand them the whole entry, including what it was tested on and what happens without the odd-looking line. The alternative is a colleague who tries it once, gets a mediocre result, and concludes the tools are overrated.

Why this is the durable part

Models change and products change. What does not go stale is the accumulated knowledge of what your work actually requires: that lead times get invented unless forbidden, that your director wants the ask before the detail, that the export uses totalAmount and not total_amount.

That knowledge lives in your file, alongside the example set and the ten test items. Those three files are what this course is actually trying to leave you with. Everything else is explanation.

Start it with three

Do not plan the library. Open a file this week and put three prompts in it: the one you used today that worked, the one you keep retyping, and the one that took you longest to get right.

Three entries is enough to be worth opening, which is the only threshold that matters. A library nobody opens does not exist, and the way people end up with one they open is by starting far too small rather than by designing a structure first.

Add an entry whenever you catch yourself thinking I worked this out before. That thought is the signal, and it is reliable.