It works, but something’s a little off: getting an Agent to make it better

The invitation page opens. You look it over: it’s pretty good overall, except the meeting time is buried so deep that finding it feels like a reading comprehension test.

You could just tell the Agent to “make it better.” But it might get busy changing colors, rearranging the layout, and adding decorations, and in the end the time is still hiding in the same spot.

This is when one clear observation helps more than one big adjective.

Say what you see

Try breaking “it feels off” into three sentences: what’s wrong, what you want instead, and what you’re already happy with.

For example: “The meeting time is at the bottom of the page, and first-time readers can easily miss it. Please move it under the heading and make the text bigger. Keep everything else as is for now.”

That way, it knows what to focus on, and that it doesn’t need to redecorate the whole page. You don’t need to add “use such-and-such CSS property” unless that’s genuinely something you want.

Make one small change

Go back to the same conversation where you made the webpage (that conversation is called a Session), and send this request.

Please edit the existing weekend-walk/index.html. Don't create a new page.
Move "Saturday at 3 p.m." and the meeting place to just below the heading, so people see them right away.Make the event description a bit more relaxed, like you're inviting a friend, and skip the over-the-top slogans.Keep the route, the packing list, the "Sample event" label, and the existing button jump.
After editing, make sure nothing overflows sideways at phone width, and save the file.Tell me exactly what you changed, and which effects I still need to check by opening the page.

When it’s done, look at the current file again. If the file preview still shows the old content, click the refresh button on the preview to reload it. If what you have open is a site that’s already been prepared, you need to click “Update preview” to generate a new version; refreshing the old site won’t bring in the latest files. We’ll dig into the difference between these two ways of viewing in the next chapter.

Then check only against the request you just made: Is the time easier to find? Is the location right nearby? Does the button still work? This kind of check is far less stressful than hunting aimlessly for flaws.

Point right at it

Some things are hard to describe in words, like “the card on the right” or “the second subheading.” In those cases, you can point to them right on the page.

Open the preview of index.html on the right side of the conversation, click “Annotate” in the toolbar, then click the element you want to change. It gets a number, and a small box next to it lets you jot down a note, like “Swap this for something more relaxed.” You can mark several spots in a row. When you’re done, the input box shows how many annotations you’ve added, such as “2 annotations.” Add one overall request, and send it all together.

The Agent gets more than your words: it also receives the position and content of each marked element, so you don’t have to struggle to explain “which part.” You can also select some text in the Agent’s reply or in a text file and use “Add to conversation” to bring it into your next message.

When it asks you something, it isn’t necessarily stuck

“Do you want the page to feel more lively, or more calm?” This is asking you to decide on style. Just reply in the input box. You can pick one, or give it something concrete to aim for: “Like an invite from a friend, not a mall sale flyer.”

A permission request is a different story. The screen shows “Requires initiator approval,” and only the person who started this task can decide. Look at what it’s about to do and where, then decide whether to allow it. If you’re unsure, choose “Cancel this request.” Style questions need your taste; permission questions need you to understand the action. Don’t mix up the two kinds of answers.

If what it wants to do clearly goes beyond this invitation page, rein the goal back in first. For example, you don’t need a real sign-up system right now, so there’s no reason to connect a service that collects personal information.

When something goes wrong, describe the scene

“It doesn’t work” only tells the Agent you’re unhappy, not where the problem is. Try this instead: “I opened the saved index.html, and clicking the button doesn’t scroll. The rest of the page is fine.”

If there’s an error message, include the relevant text. Say which step you took, what actually happened, and what you expected. Unrelated private details, like account info or passwords, don’t need to go in.

Sometimes you just opened an old version; only sometimes does the file really have a problem. Confirming which file and which preview you’re looking at saves a lot of back-and-forth.

Stop when you need to, and pick it back up

If you see it heading the wrong way, click “Stop current task” in the input box. The person who started the task and the Workspace owner can stop it. Once it’s stopped, a note at the end of the reply says this round was canceled.

Changes that were already saved aren’t automatically undone when you stop. First look at where the current file stands, then decide what’s next:

• If the direction is fine and you just want it to keep going, click “Continue task” while the input box is empty, and it’ll pick up from where it left off.
• If the direction needs adjusting, just write your new request and send it, like “Leave the colors alone for now and only deal with the meeting time.”

While it’s busy, you can also send your next request to the queue. If it’s urgent, use “Run next” in the queue: the Agent stops the current task and handles this one first.

After this small change, you’ve picked up a really useful habit: observe first, then make a specific request, and finally go back to the result to check it. You can use it for intro pages, checklists, and small tools down the road.