When the answer misses the point

What does a missed request look like?

Some AI replies are wrong in an obvious way: they crash, or they show nothing. A missed request is harder to see. The reply is tidy, the code runs, the message sounds pleased with itself, and it does something other than what you asked. Sometimes it adds what you wanted plus a redesign. Sometimes it replaces your work with its own idea of a better page.

In the practice area, ask for a Book a repair link under the bike shop’s price table. The reply turns your table into cards, adds a Book now button that calls an openModal() function that doesn’t exist, and never links to your booking form. The message calls this a refresh. If you only glanced at the Preview, you might think it was progress.

Why do assistants miss the point?

An AI assistant tends to produce the kind of answer it has seen most often. Requests about booking are often followed, in its training text, by pop-ups and booking services, so it drifts that way. It also tends to be generous: asked for one change, it “improves” things nearby, because many answers it learned from did.

Your request may also be less clear than you think. If it doesn’t say “keep my table as it is”, the assistant has no reason to keep it. Even with that sentence, it can still drift. Clear requests reduce misses. They don’t stop all of them.

Watch for the signs in the reply’s message, too. Words like “refresh”, “tidied up”, “while I was at it” or “modernised” usually mean the assistant changed more than you asked. A reply that never mentions the exact thing you requested, such as the link to the form, is another warning. Read those messages as a list of changes to check, never as proof that the job is done.

How do you compare the reply with what you asked?

Before you press **Use this code**, read your own request again and list what it asked for. Then check each item against the reply, and check what else changed. For the bike shop that means three questions: is there a link under the table, does it go to the form, and is everything else the same?

The **Review** view is built for this. After you use a reply, it compares the assistant’s code with yours and highlights every change you make. Comparing against your starting page in your own head is harder, so keep the original code open, or save it first, before you use a reply that might rewrite it.

• Is the thing I asked for there?
• Does it work when I click it?
• What changed that I didn’t ask to change?
• Is anything missing that was there before?

How do you ask again?

Say what you asked for and what you got instead, in plain words, then repeat the request with a firm limit. Blaming or repeating the original request word for word usually gets a similar miss. Naming the gap tells the assistant exactly where it went wrong.

Send this in the practice area. The **Prompt check** view ticks What was missed, and the reply goes back to your page with one link added. The phrase “only add the link” is there for a reason too: it tells the assistant to leave everything else alone.

That missed what I asked. I wanted one link under my existing price table that jumps down to my booking form, but you replaced my table with cards and added a Book now button that opens a pop-up instead. Please keep my table and form exactly as they are and only add the link.

Is the second answer right?

Better, and still worth testing. Use the code and click Book a repair in the Preview. Nothing happens, because the link says href="#booking" and the form says id="book". An in-page link jumps to the element whose id matches the text after the #, letter for letter. One mismatched word and the link goes nowhere.

Fix it by changing either side so they match, then click again. The page jumps to the booking form. This is the pattern for every reply in this course: compare, re-ask if needed, then test what you get.

<a class="book-link" href="#book">Book a repair</a>
…
<form id="book">

When should you start again instead of re-asking?

If a second correction still misses, the chat may be stuck on its own idea. Long chats carry the earlier wrong answers forward, and the assistant keeps building on them. Starting a fresh chat with your current code and a clear request is often faster than a third or fourth correction.

You can also do the change yourself. Adding one link is a single line of HTML. If the assistant keeps missing a small change, it can be quicker to write it and ask the assistant only to check it.

How do you prevent misses in the first place?

Ask for one change at a time and name what must stay the same. Paste the part of the page you are changing so the assistant works on your code. Asking it to list the changes it made, at the end of its reply, also helps: a list that says “replaced table with cards” is easier to catch than the code itself.

Save a copy of your page before using any reply that might rewrite it. Later in the course you will use Git for this, which makes undoing a missed request a single command.

Only add the link. Keep everything else on my page exactly as it is, and list every change you made at the end.

How is this pretend assistant different from a real one?

The pretend assistant misses the same way every time and always gets it right on the second try, so the lesson can check your steps. A real AI tool might miss in a different way, or get it right first time, or miss again after your correction. It may also argue that its change is better.

It may even be right that cards look better on phones. That is still your decision, and it should come from you asking for it. Keep the request as the measure: the reply either does what you asked or it doesn’t, and you decide what goes on your page.

Field notes

• A tidy reply that runs can still do something other than what you asked.
• Check each part of your request against the reply, and look for anything else that changed.
• When you ask again, say what you wanted, what came back, and add a firm limit.
• An in-page link only works when the text after # matches the target's id exactly.

Easy to get wrong

A polished reply can remove content, add features that don’t work and skip the thing you needed. Looking better is a different test from doing what you asked.

The same request tends to get the same kind of miss. Naming the gap tells the assistant what to change.

Assistants that “tidy up” often drop small things like a phone link or a footer, and their summary rarely says so.

Questions people ask

Is it rude to tell the assistant it got it wrong?

No. Clear correction is the most useful thing you can give it. Be specific and calm, as you would with a colleague.

What if I like some of the extra changes?

Keep them on purpose, one at a time, after you have checked them. Ask for them separately so each one is a decision you made.

Why did it say it did what I asked when it didn’t?

The message describes what the assistant intended, in confident language. It doesn’t check its own code against your request. That is your job.

How do I find what changed in a long reply?

Use a comparison view like Review here, or a diff in your editor or Git. Reading two long files side by side by eye is slow and error-prone.