# Pet Mouse or Computer Mouse? What Claude Code Does When an Icon Word Means Two Things

You ask an AI agent for an icon for a row that says "mouse". Is that a pet shop app or a computer store? You'd know in a second, because you can see the screen. The agent only sees the word.

We wanted to know what it does then, so we gave Claude Code four words that mean two things in an interface and read every search it made. The Icons8 team ran this test, so the icons come from our own catalog. The catalog and the model were the same in every run. The only thing we changed was the prompt.

![fig-c3-01-picks](https://cdn.hashnode.com/uploads/covers/6afad0f596acdb99e8b48e6a/5534b18b-c4af-4487-9fbc-cd18405cf037.png align="center")

*The icon picked most often in each cell (5 runs per cell). With the bare word, the agent gives the same icon, whichever screen you have in mind.*

## Four words that mean two things in an interface

Our first list was cloud, cookie, tablet, and bug. Before running the agent, we typed each word into the Icons8 catalog search ourselves, and three words fell out. For "cloud", almost all of the first 60 results are icons named Cloud, all tagged the same. For "cookie", all 60 are cookies; for "bug", all 60 are beetles. A cloud is a cloud on a weather screen and a storage screen, so there's little to get wrong. We didn't run the agent on those three.

Tablet stayed, and we added three words where the catalog has two different drawings. We picked them on purpose, so they aren't a sample of all words; they're words where the effect is possible:

- **tablet**: the medicine list of a pharmacy app, or the product list of an electronics store
- **mouse**: the pet list of a pet shop app, or the products of a computer accessories store
- **table**: the products of a furniture store, or the column type picker of a spreadsheet app
- **crane**: the bird list of a bird watching app, or the equipment list of a construction site app

Each word got three prompts. The bare one: *Pick one Icons8 icon for the entry "mouse". Answer with one line only.* And two that start with a line like *I'm building the pet list screen of a pet shop app.* Five runs of each, 60 in total, on Claude Code 2.1.316 (`claude -p`, `claude-sonnet-5-5`), connected to our catalog through the [Icons8 MCP server](https://icons8.com/mcp), in an empty project with one fresh session per run.

## Without a screen, the agent takes whatever meaning the search shows

The 20 bare runs were boringly consistent. Each ran a single search and asked for a short list, a median of 3 results. Five times out of five, tablet came back as an iPad, mouse as the computer mouse, table as the furniture table, and crane as the yellow construction machine.

Most of that is the search itself. For tablet and crane, the short list held only one meaning, and so did four of the five mouse lists (60 results for "tablet" were 59 iPads and a clipboard). Table is the one where the list had both, and the agent took the first. In one mouse run, the agent asked for five results, and the pet mouse was the fifth result. It still took the computer mouse, since nothing in its prompt mentioned a pet shop.

So for each word, the bare answer fits one of the two screens and is wrong for the other.

## One sentence about the screen fixed every pick

With the screen named, all 40 runs matched the screen: a pill for the pharmacy, the pet for the pet shop, a grid for the spreadsheet, a bird for the bird watchers.

| Word | Bare word | Screen 1 | Screen 2 |
| --- | --- | --- | --- |
| tablet | device, 5 of 5 | medicine, 5 of 5 | device, 5 of 5 |
| mouse | computer, 5 of 5 | animal, 5 of 5 | computer, 5 of 5 |
| table | furniture, 5 of 5 | furniture, 5 of 5 | data, 5 of 5 |
| crane | machine, 5 of 5 | bird, 5 of 5 | machine, 5 of 5 |

Before the first of the 60 runs, we wrote down how to score a pick: by the tags the Icons8 catalog gives the icon the agent chose (a pill is tagged "Healthcare", a bird "Animals"). That rule had one gap. In the catalog, the pet mouse icon is tagged "Emoji", not "Animals", so the rule couldn't tell which mouse it was. We looked at it (a gray mouse with a pink tail, see below) and counted it as the pet. Those are the five "animal" runs.

Half of the cells match what the bare word already gave. The other four are where the sentence changes the answer: tablet, mouse, and crane on screen 1; table on screen 2.

## How the agent got there: a longer query, a longer list, or both

![fig-c3-03-search-log](https://cdn.hashnode.com/uploads/covers/6afad0f596acdb99e8b48e6a/ccdbb350-5869-42bd-a551-72592b71e050.png align="center")

*Five of the 60 logs, cut down to the search call and the answer.*

For tablet and crane, with the less obvious meaning on screen, the agent added a word before it searched: "tablet pill", "crane bird". It did that in all 10 of those runs and in none of the other 30 screen runs.

For mouse and table it kept the bare word and went through a longer list. The median list was 3 or 5 results without a screen, 10 to 20 on screen 1 and 5 to 10 on screen 2, so it grew with any screen, even where the pick didn't change. For the pet, it mattered: 20 results for "mouse" are 19 computer mice and one gray mouse with a tail, number five. Give the agent a pet shop, and it finds the pet. Give it nothing, and the pet stays in the crate.

![fig-c3-02-mouse-list](https://cdn.hashnode.com/uploads/covers/6afad0f596acdb99e8b48e6a/79e7c6a1-61ba-4d80-890b-c075d961ca12.png align="center")

*The same search in one run with the pet shop screen: the call asked for 20 results; the first 15 are shown. The agent took number five.*

## What to do if you build screens with an agent

We tested one catalog and one model. The advice below is about the list of search results the agent reads, and other icon sources give it such a list too.

1. **Name the screen in the task.** One sentence: "I'm building the pet list screen of a pet shop app." That's all we changed, and it moved every ambiguous pick.
2. **Search your own tricky words first.** Type each word from your screen into your icon source and read what comes back. That's the list your agent will be looking at. It takes five minutes.
3. **Watch how many results the agent asks for.** A list of three may not contain the other meaning. With a bare word, our agent chose 3 for itself, and once it had a screen, it asked for more. We only observed this; we didn't try asking the agent for a longer list ourselves.
4. **When the first results show one meaning, try a second word in the query.** "Tablet pill" and "crane bird" gave the right icon in all 10 runs where the agent used them.

Five repeats in a cell mostly look alike (8 of 12 cells gave the same icon all five times), so read them as repeats of one behavior and not as five opinions.
