Who can work with Skills?
🧑🧑🧒🧒 Available on any Air workspace
🔑 Any user with edit permissions on Air
A skill is a saved instruction set — a prompt that tells the AI exactly what to do when someone invokes your slash command. Think of it as a recipe: the user provides the ingredients (their images), and your skill provides the method. Writing a good skill is less about being clever and more about being precise. The AI will do exactly what you say, including the unintended parts.
The anatomy of a skill prompt
Every good skill prompt has four parts, in this order:
Goal statement — One or two sentences describing what the skill does and what must stay the same. This is the most important part. Read at the beginning, it sets the intent for everything that follows.
Input check — Before doing anything, confirm the right assets were provided. If something is missing, ask for it clearly.
Steps — The actual instructions, in order. Keep these short and concrete.
Output — What to deliver and in what format.
Configure your skill
When you create a new skill, you will set up its basic details and provide the instructions that power the AI workflow. Fill out the name, slash command, description, and source asset type, then paste your prompt into the Instructions field.
To get the best results from your custom skills, follow these best practices when writing your instructions:
Put preservation in the goal
If something must not change (dimensions, content, colors, composition), say so in the goal line. "Relight the photo to look like daylight. — not buried in a constraint at the bottom. The AI reads the goal first and uses it to frame everything else.
❌ Goal: Relight the photo to look like daylight.
✅ Goal: Relight the photo to look like daylight. No other changes — composition, content, subject, or dimensions remain identical.
Less is more
More instructions doesn't always mean better results. A prompt that tries to cover every edge case often produces worse outputs because the AI has too much to hold at once. Start minimal. Add a line only when you've actually seen a problem it would fix.
🤔 If you find yourself writing a fifth or sixth step, ask yourself: is this actually necessary, or am I speculating about what might go wrong?
Be specific about what you do not want
The AI finds the nearest valid interpretation of your instructions. If you only describe what you want, it will fill gaps with something adjacent — which may not be exactly what you had in mind. Whenever the AI could reasonably pick something close-but-wrong, close the gap explicitly.
❌ Place the product on a wooden shelf with a white background.
✅ Place the product on a single flat wooden shelf plank against a plain white wall. The wall is completely white — no texture, no paneling, no cabinetry. The only wood visible is the horizontal shelf surface.
Use the edit operation for photos
There are two ways the AI can process an image:
Edit: takes your actual photo as pixel input and modifies it
Generate: creates a new image based on a text description
For skills where the real photo must be preserved (lighting edits, watermarks, mockups), always specify edit explicitly. Otherwise the AI may regenerate the image from scratch — which loses all the original detail.
Use the edit operation — take the provided image as direct pixel input. Do not use the generate operation.
Always include an input check
Every skill should start by confirming it has what it needs. If a required image is missing, the skill should ask the user to provide it — as opposed to failing silently or producing garbage output.
Step 1 — Input check: If no image is provided, ask: "Please share the photo you'd like to edit." For specific requirements (landscape orientation, a face present, a logo), check for those too and tell the user what to fix.
Support multiple images explicitly
If you want your skill to work on batches, say so. Otherwise, the AI may only process the first image.
If multiple images are provided, process each one independently and deliver one output per image.
Try concrete analogies for tricky operations
Abstract instructions often misfire. If you're describing something visual or spatial, a concrete analogy beats a technical description.
❌ Apply a circular mask with transparent corners.
✅ Cut the square into a circular shape: keep all pixels that fall inside the circle and make the corner areas fully transparent. Think of it like cutting a coin out of a square piece of paper — you keep the circle, discard the corners.
Set sensible defaults
If your skill has configurable settings (position, size, opacity), pick a good default and let users override it inline. This keeps the skill usable for the most people without extra steps, while giving power users control.
Default: bottom-right corner, 50% opacity, 15% image width. If the user specifies a different position, opacity, or size in their message, use that instead.
Test your skill
Your skill will be used on images you have never seen. Before publishing, test your skill with
different aspect ratios
inputs that are slightly wrong
very simple and very complex subjects
images with no face / multiple faces / background people
A skill that fails gracefully by asking the user a clear question is always better than one that produces a confusing output.
Know what Skills can't do reliably (yet)
Keep in mind that some operations are inherently inconsistent and are worth the time in skill-building to set expectations for.
360-degree rotations from a single image realistic: The AI has to invent unseen sides, which leads to hallucinated details. Results improve significantly with 4+ angle photos. Always include a disclaimer in the output.
Realistic text rendering on unseen surfaces: The AI struggles to accurately reproduce text on unseen surfaces. If your skill involves rotating or wrapping text around a 3D object, expect inconsistency.
Very subtle edits: If the effect you want is small (a slight warmth shift, a minor crop), the AI may judge its own output as "already good enough" and make minimal changes. Be explicit about the degree of change you want.
Summary: The Checklist for a Publishing Great Skill
☑️ You've included a goal statement that states the objective and what must NOT change
☑️ Your skill checks for required inputs and asks clearly if something's missing
☑️ Skill Instructions are as short as possible while still being specific
☑️ You've described what you don't want, not just what you do
☑️ If the skill edits a real photo, you've specified the edit operation
☑️ You've tested on at least 5 different assets
☑️ You've tested what happens when the user provides the wrong input
☑️ Output format is clearly specified (PNG, transparent background, dimensions, etc.)
Now that you know how to write a great skill, learn how to manage and access them in your workspace. Check out our guide on Explore Skills to dive deeper into command-driven workflows.