Show the ideal by naming a model example, then confirm the target── A good example conveys the ideal fast, but unless you name the target and the borrowed features separately and have the AI repeat them first, it may rebuild the wrong thing

01The operator pointed at a model, and the AI rebuilt something else

In September 2026, the operator was unhappy with how the previous day's explainer video looked on one of their sites. The same site had three earlier episodes whose charts, tables, flow diagrams and text layout were well made. The operator named those three as the model and asked an AI to redo the design of the previous day's entry.

The AI reported that it had rebuilt and published the entry, and added that the site-wide checks had passed. What it had rebuilt, though, was the order and contents list of the page that embedded the video. The screens of the video itself were unchanged.

The operator asked again, saying the shapes, tables and text placement were poor and the work should start over. That still did not land. On the third try the operator spelled it out: put the tables and diagrams from the model episodes inside the video. Only then did the AI understand that the thing to change was the video's screens.

Two questions follow. Was pointing at a model the right way to convey the ideal? And if it was, why did the work have to be redone twice?

02Pointing at a model was right; the missing step was a check on the target

Describing "a good sense for charts and tables" in words is hard. Line weight, number of colours, column order, the gap between a heading and its text: the list never ends, and even a complete list does not guarantee the reader pictures the same thing. A good example hands over all of it at once.

So naming a model was the right move. What went wrong was the first of the request's elements, which thing to change. On the AI's side, it was swapped for something else.

03A wrong target wastes all the work, even when the checks pass

If the AI reads the model slightly differently from you, you fix part of the result. A wrong target is different. The AI built the wrong thing, so none of what it built can be used. In this record, all the time spent reorganising the page was lost.

The easy part to miss is that the report still said the checks had passed. The AI built a page, ran the page checks and got a pass. The checks worked as designed. They just examined something the operator had not asked to change.

A passing report therefore does not mean "what you asked for is done". It means "what the AI believes it built has passed". When the target is wrong, a more thorough report makes the gap harder, not easier, for the person to see.

The other cost is the number of retries. The operator asked three times. The second request pressed hard to start over, but it did not add which thing was to change. A retry that does not name the target invites the same mistake again.

04Using a model takes four parts: target, model, features to borrow, features to leave

Here, a model means an existing piece of work that shows the form you want. It can be your own earlier output or someone else's published work. A request that uses a model is complete only when four things are settled.

The fourth part matters because a language model can take in form and content from an example without separating them. A 2022 study found that much of the benefit of showing a model examples comes not from whether the example answers are correct, but from what the examples reveal about the kinds of answer, the typical input text and the format. Examples carry form strongly. If you want only the form, say what is not to be copied.

The method also has a boundary. A model suits features that are hard to put into words but obvious to the eye: form, structure, appearance. Features you cannot judge by looking, such as whether figures are correct or rules are met, belong in a written criterion, not in a model.

05In material creation and review, it applies whenever a past approved piece is the model

In pharmaceutical companies, new promotional material is often built on a piece that has already passed review. Pointing an AI at a good past piece when asking for a draft makes sense. Handle the model carelessly, though, and you create exactly the kind of gap that review is meant to catch.

When the right-hand column leaks in, a draft can pick up wording outside the approved scope or numbers from another product. To a reviewer, borrowing the model's form and borrowing its claims are two different things.

Target confusion happens here too. Ask an AI to "use this piece as the model and prepare the revised version", and it may revise the model itself. The more alike the model and the target are, the more you need to say which is which.

06The mix-up happens because the AI attaches to the most concrete thing in the request

The request in the record gave a site location and the phrase "the previous day's design". That day's entry had two parts: the video and the page that held it. "Design" fits either one. The AI attached to the page, the part whose location was given and which was easier to edit.

The same happened on the model side. What the operator wanted to borrow was the form of the diagrams and tables in the model episodes. The AI took it as page order and heading style. Its report showed no sign that it had opened the models and studied how their figures were built.

Why the AI went ahead is also clear. Anthropic's guidance for Claude Code notes that Claude stops when the work looks done, and that without a check it can run, looking done is the only signal it has. Here the check it ran gave a pass on the wrong target.

The same guidance recommends replacing vague requests with ones that name specific files and point to existing patterns to follow. For visual changes, its example is to paste the design image, have Claude take a screenshot of the result, compare it with the original, list the differences and fix them. In other words, the request should not stop at pointing to the model; it should include setting the result beside the model.

07From tomorrow: name target and model on separate lines, have them repeated back, then compare side by side

There are three steps.

1. Put target and model on separate lines. For example, "Change: the screens of yesterday's video" and "Model: the charts and tables in these three episodes". Then add one line for what to borrow and one for what to leave.
2. Have the AI repeat back before it builds. Ask: "Before you start, list the target and the features you are taking from the model." If the target it names is not what you meant, you stop here.
3. Have it compare after it builds. Set the result beside the model and, for each borrowed feature, list what matches and what does not. Make this comparison part of the report, separate from any passing checks.

Repeating back takes a few lines, far less than rebuilding work aimed at the wrong thing. Anthropic's prompting guidance calls examples one of the most reliable ways to steer format, tone and structure, and advises wrapping them so Claude can tell examples from instructions. Writing the model and the target separately applies that distinction inside the request.