Once you use AI most days, a pattern shows up. You keep asking for the same handful of things: the monthly supplier order, the quote follow-up, the end-of-month summary, the roster problem.
And you keep typing them from scratch. Slightly differently each time, slightly worse on the busy days, which are exactly the days you needed it to be good.
The fix has 3 names depending on where you meet it, and underneath they are one idea.
A workflow is a repeatable sequence you have written down: this, then this, then this. It can live in a note you paste. This is the concept and it is the only one you need.
A skill is that same recipe saved into the AI itself, so you can call it by name instead of pasting it.
A command is the shortcut that triggers one, often typed with a slash, like /weekly-review.
You do not need any special software or setup to get most of the value. A well-built workflow in your notes app, pasted when the job comes around, captures the bulk of it. The shift that matters is not technical. It is going from "ask the AI things" to "decide once how a job gets done, then run it".
Build one by being interviewed
Do not write the workflow yourself. You will leave out the parts you do automatically, which are the parts that matter. Get it to interview you.
I do this job every [week / month / quarter]: [describe it roughly, the way you would to a mate].
Interview me about it. Ask up to 6 questions, one at a time, about how I actually do it, what inputs I need, what good output looks like, and what usually goes wrong. Do not write anything until you have asked them.
Then write me a reusable master prompt for this job with: my standing context at the top, the steps in order, the exact format the output should take, and a line telling you what to do if I have not given you enough to work with. Write it so I can paste it in cold next month and get the same quality.
One at a time is deliberate. Ask for 6 questions at once and you get a form to fill in; ask one at a time and the third question is shaped by your second answer, which is where the useful stuff comes out.
A worked example, start to finish
A cafe owner does the same job on the last Sunday of every month: work out what to reorder, what stopped selling, and what to change on the specials board.
What she typed first: "I do a monthly stock and menu review. Interview me about it."
What it asked, in order: What data do you actually have in front of you? What decisions come out of it? What do you usually get wrong? Who else needs to know? How long do you want it to take? What does a good version look like?
What came back, trimmed:
You are helping me run my monthly stock and menu review for a 14-seat cafe. I will paste the point of sale export and my supplier invoices for the month.
Work through in this order: 1. Top 10 and bottom 10 items by units sold, with the change against last month if I have given you last month's review. 2. Anything with rising cost per unit from the invoices. Flag anything up more than 5 per cent. 3. Where 1 and 2 collide: items selling well that are getting more expensive. These are the price or portion decisions. 4. 2 suggestions for the specials board based on what is selling and what is in season. 5. The reorder list.
Rules: use only the numbers I paste, never estimate. If the export is missing something you need, say so and continue with what you have rather than guessing. Keep the whole thing under 400 words. Blunt, no encouragement.
What made it good: step 3, which she would never have written herself, because in her head "what is selling" and "what costs more" are 2 separate jobs. The interview surfaced that the interesting answer is where they overlap.
That is the general pattern. The interview finds the step you do without noticing.
Make this prompt yours
Name a job you repeat and build the interview prompt for it. Nothing you type leaves this page.
Where to keep them
A workflow nobody can find quietly stops existing, and this is how most of these die.
Pick one home and use it for all of them. One note in your phone, one folder, one document. Name them for the job, not the tool: "monthly stock review", not "AI prompt 3".
If your AI supports saved skills or projects, that is a better home again, because the workflow lives where the work happens instead of in a note you have to remember to open.
Do not systematise too early
The strong temptation after reading something like this is to go and build 8 workflows this afternoon. Resist it.
Do the job manually with AI 3 or 4 times first. The workflow you write after doing it 4 times is far better than the one you guess at, because you will have learned what you want, what the AI keeps getting wrong, and which step you always end up redoing.
A workflow built on a guess encodes the guess and then you run it every month.
Revise them when they disappoint
These are living documents. Each run teaches you something, and the fix is usually one sentence.
That output was not quite right because [what was wrong]. Rewrite the master prompt to prevent that next time, and show me only the lines you changed.
2 minutes, and the version you run next month is better than the one you ran this month. Do that 4 times and you have something genuinely valuable that nobody else has, because it is shaped entirely around how you work.
When not to build one
Not every job should be one. If you do something twice a year, writing a workflow costs more than it saves. Recurring and irritating is the bar.
They go stale like everything else. A workflow built around a supplier you no longer use, or a report format that changed, produces confident nonsense. Read the standing context at the top before you run one you have not used in a while.
They do not fix bad inputs. A monthly review workflow run over an incomplete export gives you a beautifully formatted incomplete review. The rule in the example, "say so and continue rather than guessing", is worth putting in every one you write.
Saved skills and commands vary between platforms and change often. The workflow in a note works everywhere and always, which is why it is the version to start with.
Pick the job you resent most
Pick the job you have done 3 times this quarter and resented every time. Paste the interview prompt and answer the 6 questions honestly. It takes about 10 minutes and you will not type that job from scratch again.