The case against the chat window for everyday work text
A chat window is the right tool when you do not yet know what you want: exploring, comparing, being argued with. It is the wrong tool for a transformation you already know how to describe, because it makes you re-describe it every time and moves the text twice for a result you knew in advance. For repeated work the useful unit is not a conversation, it is a saved instruction that runs on whatever you have selected and returns the result where the text already was.
Watch anyone use AI at work for an hour. The pattern is remarkably consistent: select some text, copy it, switch to a browser tab or a desktop app, paste it, type an instruction close to the one they typed yesterday, wait, read the answer, copy it, switch back, paste, then fix the formatting that did not survive the trip.
For one important piece of writing, that is fine and nobody should optimize it. The problem is that it is also what people do for the twenty second edits: shortening a Slack message, fixing the tone of an email, turning three bullets into a paragraph, translating a sentence. The ceremony is identical whether the task is worth two minutes of thought or none.
What the loop actually costs
The clock time is the smallest part of it. Thirty seconds, thirty times a day, is a quarter of an hour, which is annoying but survivable. Three other costs are larger and much harder to see.
- The context switch. Leaving the document you were writing to visit a different application, then returning, is not free the way scrolling is. You come back and spend a moment finding your place, and you were mid-sentence when you left.
- The instruction drifts.Because you retype it, you retype it differently. "Make this shorter" on Monday and "tighten this up, keep it professional" on Thursday produce different results, so your output is inconsistent for reasons that have nothing to do with the input.
- The threshold rises. This is the expensive one. When each use costs a minute of ceremony, you stop reaching for AI on the small things, which are most of the things. The tool ends up used for the tasks big enough to justify the friction, which are the tasks you were most capable of doing yourself.
Chat is a shape, and shapes have jobs
None of this is an argument that chat interfaces are bad. They are a specific shape with a specific strength: they are excellent when you do not yet know what you want.
Pressure-testing an approach, comparing three options, being told why your plan has a hole in it, working something out by describing it badly and then better: all of that needs turns. You need to see a response to react to, and the value comes from the exchange rather than from any single output. That is worth switching applications for. It is worth a dedicated window.
The mismatch is using that shape for a transformation you can already describe completely before you start. When you know exactly what you want, a conversation is an elaborate way of asking for it: you type a paragraph of instruction, get one answer, and never use the second turn the interface was built around.
The other shape: the instruction as the artifact
If the instruction is stable, it should be a thing you keep rather than a thing you retype. That single change moves everything else.
You write it once, carefully, the way you would not bother to write it in a chat box: including the edge cases, the tone, the format you want, and the discipline to return only the result and no preamble. Then it is available by name. Select text, invoke it, and the output arrives where the text was.
The consequences are larger than the mechanic suggests. The output stops drifting, because the instruction stopped drifting. The prompt becomes worth investing in, because effort spent on it pays out every time rather than once. And the threshold collapses: at one keystroke, running an action on two sentences is no longer an overreaction, so the small work gets the help that was previously reserved for the big work.
What this looks like in practice
Notice a repeat
The second time you type roughly the same instruction, that is the signal. Not the tenth. The second is when you still remember what you wanted and have not yet built a habit around retyping it.
Write it properly, once
Longer than you would type into a chat box. Say what to keep, what to cut, what format to return, and that the answer should be the result alone with no introduction.
Give it a name and a place
A name you will recognize in a list at speed. If you run it daily, give it its own keyboard shortcut and skip the list.
Run it where the text is
Select, invoke, and the result replaces the selection in the app you were already in. No window, no copy, no return trip.
Fix the prompt, not the output
When a result is wrong, the useful move is to edit the instruction rather than to correct the output by hand. You are correcting it once instead of forever.
That fourth step is the one people underestimate until they have it. The result appearing in place, rather than in a window you have to collect it from, is what removes the round trip rather than shortening it. The mechanics are in how to run your own prompts on selected text.
Where a saved action is the wrong answer
Being honest about this matters, because a tool that claims to replace conversation entirely is describing a workflow nobody has.
- One-off work you will never repeat. Saving an instruction you use once is pure overhead.
- Anything you need to iterate on. If you will react to the answer and ask again, you want turns, and a chat window is built for turns.
- Work where the input is not text you can select: a codebase, a long document you need reasoned about as a whole, a genuinely open question.
- Cases where you want to be argued with. A saved action does what it is told, which is the entire point and also the limitation.
The realistic setup is both, used for what each is good at. A handful of actions covering the transformations you repeat, and a chat window open for the thinking. What changes is that the chat window stops being where you go to shorten a paragraph.
The habit worth changing first
Do not try to build a library. It is the standard mistake, it takes an afternoon, and most of what you write will be prompts for tasks you imagined rather than tasks you do.
Save one. The transformation you have typed most often this month, written properly and given a shortcut. Use it for a week. The second and third will suggest themselves from what you catch yourself retyping, and they will be better than anything you would have guessed. If you want to see what other people ended up with, the use cases are the honest list, and the free tier is enough to find out whether the habit sticks.