# Misconception Generator

Local prototype · 2026-10-26

A small Gemini-powered experiment that turns an AI feature description into three potential user misconceptions, helping designers refine capability explanations and guidance.

## Design question

What might a reasonable user believe this AI feature can do that its actual behavior does not support?

Describe the feature and ask the model to propose likely misconceptions in the user’s voice, paired with a clarification. Treat the results as research hypotheses.

## My contribution

The feature-to-misconception interaction, role prompt, response format, and streaming presentation.

## Workflow

1. **Describe** — Provide a specific feature and its actual limits.

2. **Generate** — Request three potential mistaken expectations.

3. **Clarify** — Pair each expectation with the relevant capability boundary.

4. **Check** — Review against the real product and validate with users.

## Mechanisms

**Prompt design** — A role prompt frames the model as an AI UX researcher and requests three concise misconceptions with clarifications.

**Streaming** — Gemini output is streamed into the interface so users can read it as it arrives.

**Scope** — This is a single model call, without a tool-execution loop or multi-agent orchestration. Format requirements are prompt instructions, not schema guarantees.

## A feature description is also an expectation contract.

Authored illustration · not a captured model response or research finding

**Feature** — An image tool offers “Edit the text.”

In this fictional example, the tool regenerates the full image and does not edit a text layer.

**Expectation** — “Only the words will change.”

A user may reasonably expect the layout and other image details to remain untouched.

**Clarification** — “This edit regenerates the image.”

Explain that surrounding details may change, and retain the original for comparison.

**Validation** — Check the interpretation with real users.

Ask people what they expect to happen before the edit, then compare their expectation with the output.

The illustration explains the method. Actual suggestions require review against the product’s real behavior and evidence from users.

## Current state and next steps

The feature-description input and Gemini streaming endpoint are implemented in a local prototype.

Improve the interface and compare suggestions with observed user research before opening a hosted trial.

Model-generated suggestions are hypotheses, not research findings. No user-study validation or measured accuracy is claimed.

https://linadong.design/lab/misconception-generator