From chatting to a workflow

What separates people who are fast

Watch somebody who gets a great deal out of these tools and somebody who does not, and the difference is rarely prompting skill. It is that the fast one has stopped starting from nothing.

They have four or five tasks they do repeatedly, and for each one they have a saved instruction that already contains the context, the format, the standards and the examples. They paste the new material and get a usable answer on the first attempt. The slow one types a fresh description of the same job every time, gets a generic answer, and spends three turns steering it back.

The gap is not two minutes per task. It is that the first person’s answer is good enough to use and the second person’s is not.

What a saved instruction contains

Take the last five times you asked for the same kind of thing, and write down what you had to explain every time:

ROLE:     you are helping a two-person accountancy practice in Lagos
CONTEXT:  clients are small traders; they read English but not jargon
FORMAT:   short paragraphs, no bullet lists, under 200 words
STANDARDS: never state a tax figure without saying it must be checked
          against the current schedule; never guess a deadline
EXAMPLE:  <one reply you were happy with>
TASK:     <the new thing goes here>

The example is the part people leave out and the part that does most of the work. One instance of an answer you liked communicates more about tone, length and register than a paragraph describing it, and it is the cheapest quality improvement available anywhere in this course.

Where to keep it

Products offer projects, custom instructions, saved prompts, and named assistants. All of them are convenient and all of them live inside one company’s product.

Keep the master copy in a plain text file of your own, and paste it into whichever product’s feature you are using. Then when you switch tools, or a feature is retired, or you want the same instruction on your local model, you already have it. This takes no extra effort (you were writing it anyway) and it is the difference between a workflow you own and one you rent.

The three levels, and when to move up

Level one: a saved prompt you paste. Works everywhere, needs nothing, portable. Most people should stop here, and most never reach it.

Level two: a project or custom instruction. The product applies your context automatically. Convenient, one product only.

Level three: a script. For anything you do more than a few times a day, or over more than a handful of items, a short script with the instruction inside it removes the copying and pasting entirely. The threshold is roughly: if you are doing the same thing twenty times, write the loop.

The habit worth building

When an answer comes back good, ask one question: will I want this again?

If yes, spend ninety seconds saving what produced it: the instruction, the example, the correction that fixed it. That is the entire discipline. People who do this accumulate a set of tools shaped exactly to their work. People who do not have the same conversation every week for two years.

Where the instruction should live in the request

One detail that changes results more than its size suggests. Put the long stable material first and the specific task last. Reference documents, standards and examples at the top; the new item and the question at the bottom.

This helps for two separate reasons covered earlier in the course: models attend more reliably to the end of a long input, so the actual request lands where it is noticed, and prefix caching on paid APIs matches on the unchanging beginning, so the stable half is billed at a fraction. The same ordering is right whether you are paying by the token or not.

What not to systematise

Not everything should become a workflow. Exploratory thinking, one-off questions, anything where you do not yet know what good looks like: these are better done loosely, and forcing them into a template produces confidently formatted answers to the wrong question.

The rule of thumb: systematise the tasks where you already know what a good answer looks like, and stay conversational where you do not. The first is production; the second is thinking, and they benefit from different handling.