Iro AIWhat you'll be able to do
- Generate a dozen concept directions for a brief in minutes
- Write image-generation prompts that control subject, style, composition, and lighting
- Draft realistic UX and placeholder copy instead of lorem ipsum
- Turn a pile of conflicting feedback into a clear, prioritized list
- Draft a design brief or spec that a stakeholder can actually sign off on
Inside the path
A focused set of five-minute lessons. Each one ends with a hands-on exercise, not a quiz you can guess.
Ideate directions, not final art 5 min
Use AI to widen the concept space fast: moods, metaphors, and directions you can then design for real.
Image prompts that behave 6 min
The anatomy of a controllable prompt: subject, style, composition, lighting, and hard constraints.
Copy that isn't lorem ipsum 5 min
Draft realistic UX microcopy, empty states, and error messages so your mockups test true.
Synthesize the feedback pile 5 min
Turn ten conflicting Slack comments into a prioritized, de-duplicated list of actual changes.
Briefs and specs, faster 5 min
Draft a design brief with goals, audience, constraints, and success criteria in minutes.
Try a sample exercise
This is the kind of card you'd practice inside Iro: you do the thinking, then get feedback.
◆ Sample exercise · Prompt practice
You need a hero image for a fintech landing page and want to generate it with AI. Your first instinct is "a nice image of finance," which gives you a random, unusable mess every time.
Your task: Pick the image-generation prompt that gives you control over what you actually get.
- "A nice modern image about finance and technology."
- "Make me a good hero image for a fintech website."
- "A single hand holding a smartphone showing a rising line chart, minimalist 3D render, isometric composition with the phone off-center left, soft studio lighting from top-right, deep navy background with one cyan accent, generous negative space on the right for headline text, no logos or readable text."
- "Finance app, professional, high quality, 4k, trending, beautiful."
See why the detailed prompt wins
The winning prompt controls every lever that decides the output: a clear subject (a hand holding a phone with a rising chart), a defined style (minimalist 3D render), explicit composition (isometric, phone off-center left, negative space on the right for a headline), specific lighting (soft studio light from top-right), and hard constraints (navy background, one cyan accent, no logos or text). The others leave all of that to chance: "nice," "good," and quality-word soup like "4k, trending, beautiful" tell the model nothing about what you pictured. In Iro you'd rewrite a vague prompt into a controllable one and get feedback on which levers you left unspecified.
Why designers get random results from image AI
Image models aren't mind readers. They fill every gap you leave with an average guess. "A nice image of finance" leaves almost everything unspecified, so you get a different random result each run and no way to steer it. The designers who get usable output write prompts like a direction to an illustrator: what's in frame, in what style, arranged how, lit how, and what must not appear.
Control comes from five levers: subject, style, composition, lighting, and constraints. Name all five and the model has a job to do instead of a vibe to guess. Change one at a time and you can actually iterate toward what you pictured.
The highest-leverage uses for designers
- Concept ideation: a dozen directions, metaphors, and moods for a brief (raw material for real design, not final art).
- Image prompts: controllable prompts for hero art, icons, and textures, specified down to composition and lighting.
- UX & placeholder copy: realistic microcopy, empty states, and error messages so mockups test like the real thing.
- Feedback synthesis: collapse scattered, conflicting comments into a prioritized, de-duplicated change list.
- Briefs & specs: first-draft briefs with goals, audience, constraints, and success criteria.
What AI doesn't do: have taste. It explores and drafts; you judge, refine, and ship.
What does an AI-assisted design workflow actually look like?
It looks like four or five short prompts spread across a project, not one giant prompt that returns a finished screen. Here's a real sequence for redesigning the empty state of a habit-tracking app.
1. Widen before you narrow. "Here's the screen, the user, and the single action we want them to take. Give me eight distinct directions for this empty state, one line each, from practical to strange. Ideas only, no visual descriptions." You keep two and throw out six, and that ratio is the point.
2. Write the words before you draw anything. "Draft three versions of the headline and one line of body copy for direction three. Headline under twelve words, second person, no exclamation marks, and make one plain, one warm, one blunt." Copy decides layout more often than designers admit, and knowing the real string length changes how you space the screen.
3. Then the art. Only once you know what the screen says do you spend prompts on illustration, using the five levers above and changing one at a time.
4. Run a critique pass. "Act as a senior product designer reviewing this flow. Here's the copy and the layout, so name the three weakest points and what you'd test first." Treat the answer as a checklist to argue with, not a verdict.
5. Compress the feedback. After review, paste the whole thread: "De-duplicate these comments into a prioritized list. Separate what's a required change, what's personal preference, and what needs a decision from someone else."
Each step takes a couple of minutes. The win isn't speed on any single prompt. It's that you stop building things you should have thrown out at step one. Iro's Prompt Lab, part of Pro, is where you can run rewrites like these and get feedback on the version you wrote.
How do you check an AI-assisted design before it ships?
Check four things: usage rights, exact brand values, accessibility, and whether the model invented an interface pattern you don't actually have. These are the failures that cost days, and none of them are visible in a pretty JPEG.
- Usage rights. Commercial terms differ by tool and by plan, and they get revised, so check the current ones for the account that made the image. Keep a note of which assets were generated and where, because that question tends to arrive weeks later from a client or a lawyer.
- Exact values. Image models approximate. Something generated "in deep navy with a cyan accent" lands near your palette, not on it. Correct the color in your editor, or treat the generation as reference and rebuild the asset as vector so it holds up at every size.
- Accessibility. A composition that looks great full-bleed on your monitor often fails the moment real text sits on it. Check contrast against WCAG AA (4.5:1 for body text), confirm tap targets survive the crop, and make sure nothing relies on color alone to carry meaning.
- Invented patterns. Ask a model for information architecture or a flow and it will confidently describe a component your design system doesn't include, or a pattern that fights the platform's conventions. Every suggestion is a proposal to check, not an instruction to follow.
One more that's easy to miss: read generated microcopy for claims. Models write assured empty states and tooltips about features nobody built, and that language has a habit of surviving all the way into production.
The whole pass takes a few minutes. Skipping it is how a fast draft turns into a slow rework.