The quickest way to lose an hour is to ask an AI to make a small change to a document it already wrote well. You have a memo that is 90% right. One title is wrong, one figure is out of date, one sentence needs to go. So you type "update the CFO's title and remove the last line, keep everything else identical." You get back a document that is different in six places you did not ask about. This is the last-mile problem, and it is where most of the time savings from AI quietly leak away.
Why the first draft is easy and the last edit is hard
AI is brilliant at going from nothing to something. Give it a rough brief and it produces structure, tone and a credible draft in seconds. That is the hard part of writing for most people, and the machine genuinely takes the weight.
Precise edits are a different job. When you re-prompt, you are not editing, you are asking the model to regenerate the whole thing and hoping it changes only the bit you meant. It does not work like a scalpel. It works like a rewrite. Rephrase one heading and it may "improve" a paragraph you were happy with. The longer the document, the worse the odds.
There is a control problem underneath this. Editing is fundamentally about keeping almost everything the same and changing one small thing. Prompting is fundamentally about generating something new. Using the second to do the first is like repainting an entire wall because you scuffed one corner.
What the last mile actually looks like
The final stretch of any document is rarely creative. It is a short list of small, exact fixes:
- A person's name or job title.
- A number, a date, a price.
- A sentence that reads well but is no longer true.
- A heading that is one word too long.
- A stray bullet that belongs somewhere else.
None of these need intelligence. They need precision. You know exactly what the change is; you just want to make it without touching anything else. Handing that to a system whose whole purpose is generating new text is the wrong instinct, even though it is the convenient one.
Edit the document, not the prompt
The better model is boring and obvious once you see it: treat the AI output as a finished document and edit it directly. Not the raw code, the rendered page. The thing that looks like the memo. You click on the wrong title, type the right one, and move on. Everything else stays exactly as it was, because you did not ask anyone to regenerate it.
This is what a word processor gets right and what re-prompting gets wrong. In a Word document you would never delete the file and redescribe it to fix a typo. You would click and type. The problem is that AI hands you HTML, and HTML has not historically had that click-and-type experience unless you were willing to open the tags.
That is the gap htmlEdit closes. You point, click and type directly on the rendered page, the way you edit a document, and the underlying HTML updates itself. Your files stay on your device, and there is no round trip to an AI for a change you could make yourself in three seconds. If you want the background on why these documents arrive as HTML in the first place, I wrote about that in why AI keeps giving you HTML.
A simple rule of thumb
Use the AI for what it is good at, and stop there. When the change is additive, a new section, a different tone, a fresh angle, prompt away. When the change is surgical, a fix you can describe in a few words and point to on the page, edit it directly.
The test I use: could I make this change faster by clicking on it than by explaining it? If yes, do not open the chat window. Re-prompting for tiny edits feels productive because you are typing, but you are really just outsourcing a two-second job and then proofreading the results. The last mile is short. Walk it yourself.