# The prompt you keep retyping is a workflow > Source: > Published: 2026-08-11 10:35:47.882399+00:00 # The prompt you keep retyping is a workflow Here’s a common habit. A customer email lands. You select it, copy it, switch to ChatGPT or Claude, paste it in, and type some version of an instruction you have typed before. Something like “draft a friendly reply agreeing to the timeline but pushing the start date to the 15th.” You get back something good. You copy that, switch to your mail, paste it in, edit a line, send. It works. That’s why nobody questions it. Maybe yours isn’t a customer email. It’s a bug report, a contract clause you want checked, a CV, a message from your manager asking where something stands. The details change but the general idea doesn’t: something arrives, you carry it over to the model, you carry the answer back. You have done it a hundred times, and each time you did the fetching and the pasting and the carrying. The interface you are doing all that for would have run on a DOS machine in 1985. A text box where you type and something types back, working about as well then as it does now. ## Chat is a very old interface Chat gets described as the new thing in software, and it is close to the oldest. Decades of interface work happened in between, and almost none of it shows up here. That is not automatically a criticism. Simple interfaces survive because they are general. A chat box makes no assumptions about what you might want, which is exactly what you want when you are thinking out loud or exploring something new. The problem is what the simple UI doesn’t do. Chat cannot see the message that just arrived. It doesn’t know what your team charges, what you promised this customer in March, or where the document lives. So you supply all of it, every time, by hand. The model does the writing and you do the logistics. ## The prompt is the specification If you keep typing roughly the same instruction, you are not prompting anymore. You are running a process, and the process is documented in the sentence you retype. That sentence is close to a workflow already. “When a customer asks about timeline changes, draft a friendly reply that agrees to the change and proposes a new start date, and let me approve it before it sends.” Same words, near enough. The difference is that a workflow starts when the message arrives instead of when you notice it, and it doesn’t need you to paste anything, because it is already looking at the thread. For a reply draft that’s a modest win: you save the copying. The interesting cases are the ones where the pasting is the hard part. ## Where it gets past what chat can do Take a quote request. A prospect emails asking what a project like theirs would cost. In a chat window you paste their email, then you go find the rate card, paste that too, maybe paste the last similar quote you sent so the tone matches, and then ask for a draft. Three retrievals before the useful part starts, and if a colleague does the same thing next week they might grab an older rate card and quote a different number. The version that runs on its own pulls those numbers from your [Brain](/brain/), where your pricing sits as canonical knowledge, carefully curated and updated on purpose rather than copied around. The workflow reads the request, loads that page, drafts the quote against the current rate card, and puts it in front of you to approve. You stopped being the retrieval layer, and the whole team now quotes from the same source rather than from whatever each person had open. Then take a request to update something. A client asks for the status doc to reflect what shipped last week. In chat you find the document, paste its contents in, describe the change, get a revision back, and paste it into the real file, hoping you didn’t clip a section on the way through. Work like that can be handed to an agent instead, including one running on your own machine, where it opens the actual file, makes the edit, and comes back for review. We have had it draft a blog post. A workflow can also reach the tool your document lives in over [MCP](/mcp/) and do it there. Either route matters for the same reason, and it isn’t speed: the file never gets round-tripped through a text box, so there’s no version of this where you paste back a truncated copy as you rush to get home for dinner. ## When chat is still the right answer Nonetheless, chat is often the right tool. Thinking through a problem, learning something new, getting feedback on a draft, writing something you will only write once: for any of that, a workflow is overhead you don’t need. Exploration is not repetition, and automating a one-off is its own kind of busywork. The signal isn’t that you used chat. It’s that you used it the same way, over and over. At that point you aren’t exploring anymore, you’re running a process by hand, and it belongs somewhere else. ## Try the translation Take the last prompt you retyped. Write out what triggered it, which is usually a certain kind of message from a certain kind of person. Write what you had to go and fetch before the prompt was any use, because that’s the part a workflow removes. Write what should happen with the result, and decide plainly whether you want it to send on its own or wait for you. If you’d rather work through your whole inbox than one prompt, we wrote a guide to [finding the work worth automating first](/guides/what-to-automate-first/). It covers which chains are frequent and patterned enough to hand over, and which ones need your judgment and should stay with you. You have just described a workflow in the same plain language you were already using. In this+that that description is the setup. The prompt stops being something you retype and becomes something that runs when the work shows up. ## Key takeaways - Chat is one of the oldest interfaces in computing, and its simplicity is why it survives. What it can’t do is see the work arrive, so you become the part that notices, fetches, and carries the result back. - If you type roughly the same instruction into an AI chat window repeatedly, that sentence is already a process specification. The only thing missing is something to run it when the message lands. - The bigger wins are where the fetching is the work: a quote drafted against the pricing your Brain holds, so nobody quotes from a stale rate card, or a document edited where it lives, by an agent on your own machine or through an MCP connection, instead of round-tripped through a text box. - Chat is still right for one-offs, exploration, and thinking out loud. The signal to automate is repetition, not complexity. - Translating a prompt into a workflow is mostly writing down what triggered it, what you had to fetch, and whether the result should wait for your approval.