Back to site
CMS

Built with AI

Describe the screen, correct what comes back, and let the widget you clicked tell you what it could do better - the catalogue is large so that nobody has to learn it

Not low-code. AI-first.

Low-code asked a good question - how do people build screens without waiting for developers - and answered it with a builder. The work moved from writing code to arranging components, and arranging components is still work: somebody learns the tool, somebody drags the blocks, somebody owns the result.

This is a different answer to the same question. You describe the screen and an AI agent composes it from the platform's own vocabulary. What a person contributes is the judgment - what the screen is for, who opens it, what belongs on it. What remains manual is correction, not construction.

The distinction matters because it changes who can do this. A builder needs a person who has learned the builder. This needs a person who knows what the screen is for - which is a different person, and usually the right one.

A large catalogue is a problem for people and an advantage for an AI agent

A hundred and seventy widgets is a lot to know. In a conventional page builder that is a real cost: somebody has to learn what exists, remember which block does what, discover the seventh property that turns out to be the one they needed, and keep all of it in their head between the two times a year they build a page.

That cost is what makes page builders end up in the hands of one person who "knows how it works".

Here the catalogue is read by the AI agent, not by you. Every widget declares what it is, what it accepts and what it does, so it works from a description of what is available instead of from a memory of it. The width of the catalogue becomes an advantage instead of a curriculum.

Describe the page

Say what the screen is for and who opens it:

A page for department heads. At the top, this quarter's budget: allocated, committed, spent, as three tiles. Under it, the requests waiting for their approval, as a list they can act on. Then a chart of spend by category for the year, and next to it the three largest requests this month.

What comes back is the page: the widgets chosen, bound to the right data, arranged, with the visibility rules that make it right for a department head and not for everybody.

Then you look at it and change what is wrong - by moving things yourself, or by saying what should differ. Both edit the same page.

Select a widget, and it tells you what it could do

This is the part that removes the rest of the learning curve.

Click any widget and it offers what is worth considering for that widget, in that context: the number is negative, so highlight it in red; this list is long, so it wants pagination or a filter; this table has a column people will sort by, so make it sortable; these tiles are inconsistent with the ones above; this text is doing the job of a heading, so make it one.

You did not have to know that the widget had a conditional-formatting property, or what it is called. You clicked the thing you were looking at, and it told you.

And you can say it instead. Select a block and describe the change - "put the totals on the right", "make the status a coloured chip", "this should only show when the request is overdue" - and the AI agent edits the same properties you would have edited in the panel. Point when pointing is faster; describe when describing is faster.

Why this is safe to let it do

Because a page is a description, not code.

  • What the AI agent produces is the same thing a person produces. There is no AI-generated page that a human editor cannot then open - one format, one editor, one validator.
  • It cannot produce an invalid page. Every widget's properties are declared and checked; a property that does not exist or a value of the wrong type is refused as it is written rather than appearing as an empty area at nine in the morning.
  • It cannot show data somebody may not see. Binding a widget does not grant access to anything. The platform's access rules decide what comes back, for the person looking, whatever the page asks for.
  • And it is versioned with the application. A page you changed ships with the version, diffs against the last one, and rolls back with it. See Applications.

Where this leaves people

Doing the part that needs judgment: deciding what a screen is for, who it serves, what belongs above the fold and what does not belong at all. Those are questions about your business, and no catalogue of widgets answers them.

The arranging - which was most of the work and none of the value - is what has moved.

Next

Under the hood - what a page is, and where it gets its data.