@aaapinen

building https://nitter.cf/t.co/UCeI9NTYjE so design engineers can edit the running product. writing about ai tools, code-driven design, and shipping.

San Francisco
Joined May 2024
silt's design mode edits the interface your users actually live in. not a mock, not a copy of it. the running thing. point at any element and the rulers come up. start scrubbing and you see the vertical and horizontal alignment as you drag, so getting something level is one gesture instead of a guess. the old way was describing it to an agent and waiting. write a paragraph, wait, then find out it edited the wrong component. here you point at the thing first, so it always knows exactly which component you mean. that's the difference. micro-edits that used to cost minutes each now cost seconds, and those are exactly the edits that take an interface from sloppy to intentional. filmed w/ @shotbaseapp
1
525
design w parallel agents on a canvas that knowa your system, in code. a project in silt is your repo - existing or new. worktrees let your agents n teammates explore multiple directions at once without overwriting each other's work. everything is code, everything is interactive. toggle motion on and you can actually feel the designs on the canvas. bridging design and code. once and for all. v2 of @trysilt drops soon. dm me and i'll keep you in the loop while we build - early list's still open.
9
what if your ai's thinking state was a fun drawing?
1
62
talked to a designer today who wants to leave figma, but still needs the canvas to compare ideas, comment and annotate. generation is not the missing piece. a shared place to make decisions without drift is.
59
we're pivoting. thank you to the early users - your feedback got us here. we're packaging what we have into a mac app, and fully embracing what it's becoming: the place where you orchestrate design agents in parallel. and the canvas? everything is the real thing. interactive components with motion. just how you like it.
2
2
113
what if you could run parallel design agents like conductor - each on its own branch, each on its own canvas, designing multiple things at once, all controlled from one place. design is rooted in code - not just html and css, but js and the motion libraries you actually use. 0 drift from code to canvas and back, because everything is code. you're not designing a picture of the product. you're designing the product.
1
57
local prototypes are easy to make and weirdly hard to share. why can't the exact thing on localhost become a live collaborative canvas in one click? not a screenshot. not a rebuild. the same thing.
1
3
94
a design system for agents can't be a figma page and a markdown file. it needs typed tokens, a tiny menu of valid decisions, and lint that rejects the rest. if arbitrary values are the easy path, the agent takes it.
2
60
vibe coding has a design drift problem. the agent uses your tokens until it doesn't. then 14px and #f3f4f6 quietly move in. this isn't a better prompt problem. off-system choices should be impossible to write.
2
77
i asked 15 designers how they hand context from one coding agent to another. most don't. they copy the brief into a new chat, explain every decision again, then fix the same misunderstandings. how does your team hand this off?
1
2
117
love gathering inspo from x on the canvas and exporting that context to my ide. can't wait to show the shipped result
2
2
337
oh btw: dm me if you want my extension to download x videos by clicking on a overlayed button on the video!
13
designing with an agent gets interesting when the selection itself becomes the prompt. point at the running ui. say what should change. inspect the real result.
1
137
agent chat is useful until you need to point at the same thing. then you need a shared surface both humans and agents can edit.
48
share the running thing.
41
my talk prep for boost turku turned into a pile of notes about one idea: interfaces are disappearing. dumping them here bc i keep thinking about it
1
1
61
the part that scares me a bit: the generalist trap. ai lets you move fast across domains with surface-level understanding. feels like leverage until you need depth. meaning still needs a human in the loop
1
15
so the designer's job gets bigger, not smaller. fewer pixels left, more weight per pixel. the approval moment, the tone, whether you trust the thing. felix's piece:
8