AI makes it easier to produce more versions of a screen. It does not make choosing one easier. In Content Design, we have always had to decide what the text must achieve, what it can claim and whether it works in context. More drafts give us more to explore, not a lower standard.
Writing this article with AI brought that distinction into focus: a fluent draft could still drift away from my point. Keeping it on course meant making my criteria explicit. Below, I apply three authors' work to one message, then propose a framework for carrying those criteria through AI-assisted exploration, revision and acceptance.
The criteria are not new
Torrey Podmajersky, Sarah Winters and Kristina Halvorson, my content design Avengers, bring different strengths: examining the text, its evidence and the system that accepts it. AI makes none of them redundant.
Take a hypothetical transfer held for review. We know its status, but not when it will arrive, whether the funds are safe or whether the person must act. Assume the person needs to understand that status.
Read the text against its purpose
Podmajersky's Strategic Writing for UX names four editing phases: purposeful, concise, conversational and clear. Here, "We're just checking a few things" sounds conversational but hides the state. "Your transfer is under review" names it. The aim is not brevity for its own sake, but a sentence that does its job.
Check what the text is allowed to claim
In We need to talk about GenAI, Winters addresses AI-assisted practice directly: provide context, judge when to rely on the tool, build checks into reusable prompts, and check facts, bias and stereotypes. Applied here, her guidance means checking each claim against the brief, not choosing the most reassuring tone.
Our edit, with reasons
Before
The draft notice.
Your transfer is safe. We’re just checking a few things. It will arrive within 24 hours.
No evidence for the safety claim.
Name the known state.
No verified arrival time.
After
The working revision.
Your transfer is under review.
Decide whether the work can be accepted
A cleaner sentence is not a finished screen. Halvorson's content strategy quad connects editorial strategy, experience design, content structure and process design. Standards, journeys, reuse and accountability belong to that wider system, including the question "Who gets to say no?"
That changes our decision. Where will this message appear? Is the state label consistent? Must the person act, and how will they learn the outcome? Who can verify and approve those answers? Until we know, "Your transfer is under review" remains a working revision.
More versions do not lower the standard
The same candidate needs all three tests: does the text do its job, do its claims have support, and can it be accepted in its context? These are my applications of the authors' work, not a shared AI method attributed to them.
Defining "good" before generation is not new either. Podmajersky's chapter on LLM-based product experiences includes that question. Bringing it to my own drafting workflow is an analogy: her chapter concerns content generated inside a product.
If AI makes more versions available, direction must remain active across them. Criteria and reasons need to survive the next prompt. What we learn can also improve how the next task gets done.
A working framework for directing the work
I propose a short direction record: set the criteria, inspect a version, record the decision and carry useful learning forward.