Own tool - Built with AI - 2026
Prompt compiler for Figma Agent.
My own skill that turns what I want to do in Figma, said in my own words, into a precise prompt ready to paste into Figma's AI.
What I ask
"Create a disabled variant of the primary button."
The challenge
I use Figma Agent daily and the result depends on how the prompt is written. I wanted to stop writing them by hand every time.
What I did
A skill that translates the request into a structured prompt, with templates per task type and the design system context.
Result
I refined it through use, testing it with real prompts. Now most tasks come out right the first time.
01 - The problem
A vague prompt, a vague result
Figma Agent is part of my daily work. What it returns depends a lot on how you ask: which component, which variables to use and what not to touch.
Writing those instructions by hand every time was slow and inconsistent. I saw it as a small product problem with a very clear user: me.
02 - What it covers
The tasks I repeat most
- Components and variants: create or edit them without breaking what's there.
- Variables and tokens: apply them instead of loose values.
- Screens: generate or adapt them on the design system.
- Layers: reorganise and rename them with a shared convention.
- Documentation of components.
- Debugging: when a prompt doesn't deliver, review and fix it.
03 - How I built it
I define, AI compiles
What I decided
Which tasks it covers, how each prompt is structured and what design system context it needs to work well.
What AI sped up
Drafting and organising the templates, and turning each request into the final prompt in seconds.
After using it with several prompts, I tuned the templates until most tasks stopped needing corrections.
It follows the same idea as my earlier skill for turning Figma designs into WordPress: automate the repetitive to spend the time deciding.
04 - Next steps
Giving it more context
The next step is completing the design system context block: variable collections, naming conventions and base components. The more context it has, the better the prompt it returns.