When you start a review, you’ll notice a new “Write a prompt” button at the bottom of the category selection window. This lets you write your own prompt and tell AI exactly what to look for during the review.
If you'd like to reuse the prompt later, you can save it as a custom category.
You can choose to make the category visible only to yourself, or—if you're on a workspace plan—share it with other workspace members so they can use it in their own reviews.
While reviewing results, you can also update your prompt, save it as a new category, and rerun the review.
💁♀️ Only the creator of a custom category can edit or delete it.
Accessing your custom categories
To view and manage your custom categories, click the “Reviews” button in the top-right corner of your project dashboard, then select “Manage your categories”.
How to Write a Custom Review Prompt
Custom prompts let you define your team's specific design rules, so AI can review screens based on the criteria that matter most to you.
What Zeplin AI knows about your designs
Before writing a prompt, it helps to understand what data AI has access to during a review. Your prompts can reference any of these:
Components — names, whether they're linked to the styleguide, variant names
Spacing — padding (per side) and gap values
Typography — font family, size, weight, line height, and whether a text style is linked to the styleguide
Text content — the actual visible copy on your screens (for casing, spelling, grammar checks)
Colors — fills, strokes, and shadows, including color variable/token names and whether they match your styleguide
Layout — alignment and positioning of elements
Contrast — contrast ratios between text and background colors
Layer metadata — layer names and where layers sit in the hierarchy
Rules for writing a good prompt
The examples under each rule are good starting points — feel free to use them as a template and swap in your own values.
Make your rule measurable
The AI needs a concrete value or condition to check against. Without one, it has nothing to measure.
❌ "Check the spacing."
✅ "Flag any spacing value that is not a multiple of 8. Our design system uses an 8px grid."
Give a reference answer
Don't just describe the problem — tell the AI what the correct result should look like. This is the baseline it compares everything against.
❌ "Flag wrong fonts."
✅ "All text should use Inter. Flag any layer using a different font."
List your exceptions
If certain elements shouldn't be flagged, say so explicitly. AI reviews everything that matches your prompt. If there are exceptions, make sure to specify them explicitly.
❌ "Review all text casing."
✅ "Review text casing for body copy and headings. Ignore icon labels, navigation tab names, and placeholder text.”
Use your actual component and style names
If your design system has specific names for components or styles, use them. The AI can match against your exact styleguide names.
❌ "Check that primary buttons look right."
✅ "Flag any CTA that isn't using the Button / Primary component from the styleguide."
Keep each prompt focused on one thing
A prompt that covers multiple areas at once may produce less consistent results. One clear topic per prompt works much better.
❌ "Check spacing, colors, fonts, and whether components are linked."
✅ "Review all visible copy for spelling and grammar only. Flag any typos, missing words, or grammatical errors. Ignore layer names and placeholder text."
Tweak as you go
Your first prompt doesn't have to be perfect. Review the results and refine the prompt as needed by tightening the rule, adding exceptions, or narrowing the scope. Prompts improve quickly with one or two rounds of tweaking.
Tips for specific review types
Components: Mention whether you want to check linking, naming, or both. Example: "Flag any button that isn't linked to a styleguide component."
Colors: Specify if you care about fills, strokes, or both — and whether you want raw hex values flagged or just missing variable links.
Casing: Always list exceptions (proper nouns, brand names, short labels). Casing rules almost always have edge cases, and the AI will flag them without guidance.
Spacing: The more specific your grid values, the better. Listing the exact allowed values (e.g., 4, 8, 16, 24, 32) gives the AI less room for interpretation.
Contrast: Reference a specific WCAG level if you have one (e.g., AA or AAA). The AI can check contrast ratios against your required standard.