Writing a prompt someone else can reuse
The instruction that works when you type it often fails when a colleague does. Four things that make the difference, and why this is documentation.
In February this blog said the signal to convert a chat into a process is that only one person knows how. The reason is usually not secrecy. It is that their instruction carries assumptions they do not know they are making.
What goes missing when someone else types it
You write « summarise this for the customer ». It works, because you know which customer, what they already know, what you would never put in writing, and how formal your house style is. None of that is in the sentence.
A colleague types the same words about a different customer and gets something technically correct and unusable.
The four things a reusable prompt states
Who the output is for. « A purchasing manager at a construction firm who has not seen our previous correspondence. » This single addition fixes more bad output than any other change.
What must not appear. Prices not yet agreed. Delivery dates. Anything about the other project. Stating the prohibitions explicitly is more reliable than hoping they are inferred.
The shape of the answer. Three paragraphs. A table with these four columns. Maximum two hundred words. Without this, length and format drift every time and the person has to reformat, which is where the time saving goes.
What to do when unsure. « If the document does not say, write UNCLEAR rather than estimating. » The most under-used instruction in business AI use, and the one that converts a confident wrong answer into a flagged gap.
Where it lives
A shared document, with a line under each prompt saying what it is for and who to ask about it. Not in someone’s chat history.
That document is, in practice, process documentation — the first that many SMEs have ever written down for these tasks, because the AI made the implicit instruction explicit.
The unexpected benefit
Companies that do this report something consistent: writing the prompt clarified the process for the humans too.
If you cannot state who the output is for and what must not appear in it, that was already true before any AI was involved. It was simply never written down, and the new colleague was learning it by getting it wrong.