Episode 068 Raw Data

	A	B	C
1	Thanks for your interest in The Question! Sign up to participate in future questions here. This file is the raw data from Episode 068, deep dive hosted on February 12, 2026 with Ben Callahan and TJ Pitre. There are 184 answers. The question this week was: ----- Hello, system thinkers! We’re back with another episode of The Question. In our last session around AI, we explored the challenges that emerge when we rely on stochastic models to keep the very deterministic promises of a design system. This week, I’m joined by my friend TJ Pitre to understand the necessary work that needs to happen behind the scenes of a design system so that it can serve as powerful context for AI. As you know, this space is evolving rapidly. So, we want to hear from you what you’re doing to build trust in the output AI creates with your design system! Welcome to Episode 068 of The Question. Questions: How prepared is your design system to act as reliable AI context today? Have you experimented with having AI generate UI based on your design system? If yes, how do you feel about the output of your AI? If no, what’s keeping you from trying this?
2	Question 1: How prepared is your design system to act as reliable AI context today?	Question 2: Have you experimented with having AI generate UI based on your design system?	Question 3: If yes, how do you feel about the output of your AI? If no, what’s keeping you from trying this? If other, please explain!
3	5	Yes	Feeling pretty good, and we keep improving it. We’re working on transforming different forms of input into either designs or code using our DS.
4	1	Yes	I think the output can be impressive if context is setup correctly (in Figma and/or code). It also really depends on the use case, sometimes the output is 90% there, sometimes 60% there. That said this is the status NOW, there is no reason to think that in 5-10 years the output is going to be perfect thanks to the iterations the industry is going through.
5	4	Yes	I would say quite happy. Did a test: gradually feeding the knowledge to AI: first the foundation, then component library (code), then component documentation,... Each time the output get more close to production quality. Currently working with Design System AI Templates (with ready prompts) to help the organization start with ready foundation - like an application which already has DS npm package, MD file(s), ready layout (header, footer, main content)
6	0	Yes	I'm currently experimenting with this and using loads of content from TJ actually to get inspo. Our documentation is spread across different touch points and the structure of the content is not very AI-parsable. I'm realising that the content not only needs to migrate to the same repo to be centralised somewhere but also needs to be rewritten for AI...
7	1	I’m building the design system from scratch, with a strong focus on structure, semantics, and scalability from day one.	I’m currently exploring possibilities and experimenting with different tools, building the foundations from scratch, and learning what works best in practice.
8	3	Yes	Figma Make. Good for rapid prototyping but bad at production-ready code.
9	1	Yes	It isn't very good, both on Figma MCP and Figma Make. We have some success in using agents, but none that connect on the Figma side of things.
10	4	Yes	Without any additional context, the output is often pretty low quality, aka slop. However, with the right infrastructure (MCPs, Skills), the output can be indistinguishable from what an engineer would write by hand.
11	1	Yes	I only recently tried Figma Make using a Figma library as source and the results were atrocious.. I'm having a hard time understanding what AI does behind the scene to "use the Design System". Looks like in most cases it just blatantly ignores it altogether. This attempt didn't give me a lot of confidence to try again. But I'm very eager to learn how to make AI understand the DS context.
12	4	Yes	It has been a mixed bag so far. When mixed with other tools (Figma MCP) and given strict instructions to use the design system guidelines it seems to work a lot better. My worry is not everyone at my company will be as specific with it and we may end up with mixed systems or unsupported code. The output sometimes makes assumptions where I see bits of other design systems sneak in or it assumes classes it thinks I should have in my current design systems. However, it does seem to be getting much better with each additional model that comes out. Also my team has been putting in more work to improve our MCP offering to give better context.
13	3	we've got some downstream users trying this. we're focused on better setting up our libraries to better feed the ai; we're in a big parity push this quarter for this.	we've got some downstream users trying this. but for us (fintech w/ high risk barriers) Gems are not shareable currently. and there is a lot of manual work to feed the AI to make UI. and a lot of what is coming out "looks" like our system, but isn't the system. we're also just starting to test codeconnect w/ our developers and MCP server is not available to us yet.
14	2	Yes	I think there is a real possibility even with our ancient system for engineering and vertical teams to get a really healthy head start on the front end. Craft is easily overlooked with *waves arms around deadlines and requirements. This small step for us, sets us up for success for both design standards and clean code.
15	2	No	Not really sure where to start. There's so much out there, and no time to explore because we're working on bugs and priorities
16	0	No	I'm just not in a rush. I'm not fearing being "left behind" in the rat race that AI has become. I also work in a primarily product-oriented position where we aren't generating net-new UI almost ever (compared to spot enhancements, revisions, etc.). The need for generative AI based on our system just isn't really there at all. There's "leadership interest" in the topic, as leadership drinks "AI efficiency" up, but I think we're in a place where we're still establishing trust and stability, and changing gears to be AI ready would only exacerbate any issues we have, not improve them but automating their inclusion in UX initiatives.
17	1	Yes	Very small testing so far, not proven. More exploration and learning required. Individuals on our team have done some side quests to learn and unpack how AI makes its way to workflows for their particular roles, but it’s a bit fragmented. The main blocker to this work for my team is time and capacity. As a very small team, we’re marching toward our roadmaped business goals. AI workflows are coming so fast and changing so often that we haven’t had enough time to dive in collectively for what it means for our team or organization.
18	2	Yes	Not great. We need to train the system better.
19	3	Yes	We've had some issues with AI hallucinating props / other elements of a component that don't exist. We've also seen output from some of our users that wouldn't pass our design guidelines. One thing we're working on is trying to feed the MCP we're using up-to-the-minute vs. point in time documentation.
20	2	No	Too many tools, not sure where to start
21	1	No	At my company our AI tools are very limited for security reasons and there doesn't seem to be a lot of connection we can do between the AI agent and Figma. On the other hand, I feel very overwhelmed by AI and how it works. I feel like an elderly person trying to learn new technology. Trying to build out gems for AI feels as complicated as coding in my mind. At the same time I question the quality of output from AI. From what I heard, there seems to be a lot of mistakes one has to double check every time they use AI? So in that case, is AI really benefiting us and can folks identify those mistakes? Or will users believe the AI is always right and start pushing patterns that dont actually align with the design system?
22	3	No	The last DS that I worked on was for a bank with a security too high to use AI on. More generally, I'm waiting for more robust solution to do that like the storybook MCP that will give the ability of creating relavant component directly in the DS repo and integrating them in the final product.
23	3	Yes	Hallucinations!! Our AI tooling continuously thinks that it knows the API of our components but it just makes up API details based on other design systems. We're in the process of rolling out an MCP to help, but it's still an uphill battle to convince the AI to use tools!
24	3	Yes	It's pretty good for the most part. We on the design systems team haven't spent any time actually working to make our design system AI friendly but the nature of our documentation, and processes have proven to make easily understandable and readable code.
25	2	Yes	Set up was easy and simple. When prompting it was very quick and accurate when give the data from the system i.e. tokens. So producing engineering ready components I was confident. I also been looking at creating molecules from a set of atoms inside of the current design system and then to pass off to a agent to build with our repo coding patterns, again this was 90% accurate. A huge time saver and the ability to work at scale through this plugin and workflow is a game changer.
26	3	Yes	Need to provide a lot of context and guidance
27	2	Not specifically generating UI, but I have experimented with AI generating component specifications like APIs/Properties, WCAG standards, etc.	The output has been hit or miss. We struggle with access to tools and lack of connection between tools (no MCP server). Context is often injected manually and without much curation. The context window can easily become bloated and useless for moments of precision, such as component specifications. It's been great as a tool for research and auditing external design systems for component capabilities and inspiration. But again, quite imprecise. If you compare audit results to actual first-hand research, AI often drifts from precision and infers a lot of information, rather than directly referencing it from a requested source. But as a starting point to generate potential scope for a component, it's been great.
28	3	Yes	I have been successful in taking phased approach as supporting plugins and need to gradually integrate AI capabilities to manage and govern DS
29	2	Yes	The output is not consistent or reliable.
30	3	Yes	The output is hit or miss, but AI can be used to rapidly iterate using existing DS patterns or known UX patterns. The code behind the output, unless directly connected to a code repo, is often terrible and inaccessible.
31	3	Yes	I can build UI using AI and our system, but the challenge is making it consistent and scalable.
32	4	Yes	Pretty good with components, need more pattern context and composability regarding UX standards.
33	2	No	We do a lot of smaller product enhancement so haven't had the opportunity to build new pages from scratch which is where I think this feature would shine.
34	2	I'm just getting started on some of these things, as I'm not directly working on a system currently.	N/A
35	1	No	I believe more and more that we should be questioning the social acceptance of AI use. Even if you don't care at all about the ethical implications of how data models are used for training and the growing evidence of brain rot from outsourcing our critical thinking skills, the environmental impact is already horrifying. I'm not even saying that there are NO reasonable uses for AI, but shoe-horning it into every process in the name of efficiency is short sighted and frankly concerning.
36	4	Yes	It’s a good starting place for ideating new concepts and provides teams with a decent starting ground, but consistently misses the details on styling & design language guidance that are outlined in our design system, even when that context is fed in. It can provide a good starting point for discussions but it’s concerning to me when Product is using AI to “design & move quickly” on features that (beyond foregoing UX research, etc…) designs are going straight to engineering that result in havoc from a systems alignment perspective. We already have legacy UI (Backbone) classic UI (React) and Rebrand UI (Pendo guides, etc…). We don’t need MORE visual inconsistency with “AI UI”, so getting DS context fed properly and improving the quality & consistency of its output is essential. 😂
37	5	Yes	I’m satisfied. But that doesn’t mean enterprise brands would be.
38	2	In the process of building something out that AI can consume. It's been a bit rough and rocky, but getting better as AI gets better (IMO)...	A little hard to say, hard to know... Like I mentioned, it's a bit easier as AI advances. Writing design and code documentation clearly becomes more about "this is the guidebook/guardrails for AI agents as we move forward". And I'm not personally empowered to spent 100% of my time on it, so the whole "it gets better as time goes by" is because I'll put it down and come back to it and be amazed by what it can do when I get back around to it. Personally, I'm excited with what it might be able to do in the future and as we get moving forward on it, but it's a question of how to sell to higher-ups.
39	4	No	Most of our work at the moment is refinement and not net new or exploratory. Where there is exploration, it's being done by the teams who are interested in adding something new. Definitely not opposed to this, but just haven't felt the need just yet.
40	3	Yes	It’s been decent, but our evaluations of the output have not been rigorous. More work has been done with prototyping tools (Builder.io primarily) that generate code that is not destined for production. Developers have been leveraging the actual code base for context which we assume is not optimal, although our documentation in Storybook is pretty comprehensive.
41	2	No	For now, it still seems to create too many issues so detecting and fixing those takes longer than the time saving created by AI
42	3	No	The main barrier right now is our organization's strict IT/data/software security policies. We are a healthcare institution so we have strict policies in place to protect patient privacy and data. Every technology decision (even just using Slack, for example) can take awhile to vet. Our clinical practice is, and has been, using AI pretty judiciously for awhile now, but OUR side of the house (UX, design, etc) still has some red tape to get through before we can begin experimenting in-earnest. We're getting there though.
43	0	No	time
44	3	Yes	The product designers I work with have been experimenting with different AI prototyping tools. They’ve been mainly using Figma Make. It’s not very efficient, and the price is about to skyrocket. You can hook up your Figma design system library, and it works okay. One of the other designers started adding a Make file that has our brand and components in it, as something to start with. It’s working, but it’s a whole extra spot to maintain. Because of the cost of Figma Make, we have been exploring using Lovable. It’s working much better and more efficiently than Make. You don’t have to prompt it nearly as much to get something workable. You can integrate your Storybook into Lovable, which sounds great, but not at our current account level. We would need to upgrade. So that’s where we are now, but I’m hoping we can get access to this feature. I really want to normalize everyone using the design system. I am curious to hear about others’ experience with this.
45	3	Yes	Mainly that we still need to find a method that is reliably repeatable. Additionally, using the right model for the right output. Attempts at one-shotting from idea to final UI can become expensive quickly.
46	2	No	Design System isnt finished :(
47	3	We’re starting to rapid prototype using our DS	Working on getting out Storybook instance upgraded to support using the glorious MCP servers that are dropping.
48	2	Yes	Takes too long to set up context. Our company's DS is integrated with our core company repo, not standalone, which makes iterating and testing difficult.
49	3	No	We have used AI to build content mockups and its pretty good. I just haven't thought to use it for UI exploration yet, but that is a really good idea. When we are ready to branch out I'll be using it!
50	1	Yes	The little that I have tried has been lack luster. I understand that the process of getting this right will take some time, unfortunately every time I do go to kickoff the investment in some infrastructure around this the resources I would use have changed, It feels like the ground is always moving. It also feels like internally on my team there is a a range of enthusiasm to skepticism around what level of investment to make. It feels having strong cross functional support is essential but engineering and design have differing priorities and timelines for things.
51	3	Yes	Figma Make - some of our code props and Figma variables differ in name. Some things in code have been impossible to mimic accurately in Figma and needs text explanation. Still have not received Figma Slots access from announcement at Schema Fall 2025. Need to explore Storybook MCP more. All of our docs are in Storybook so we are up for this idea.
52	0	Yes	I had Claude write guidelines for Figma Make using references from the codebase. It's 90% of the way there, but makes contrast errors, readability issues, the basics that I assumed will be the easiest to apply standards to. Currently problem solving fixing these rules.
53	3	No	Our team has been focused on other initiatives and hasn't allocated time to experiment with AI UI generation tools
54	3	No	Our security team blocked the use of Figma Make and other AI tooling, but I intentionally architected components and named props/values based on the HTML and ARIA specs. I'm curious if that will cause the AI to generate better UI and/or components.
55	0	No	The current AI implementation of our DS into designs like Figma make have been inconsistent. So to build making components from AI have hindered me.
56	3	Yes	It is a hit or a miss. It is getting better. We have been experimenting for more than a year now and it feels we are closer than ever now for this to work.
57	1	Yes	Still not entirely reliable. We use Figma Make and it makes up a lot of things if it doesn't have the right context.
58	1	Yes	Just Figma Make during its early stages; it was really terrible. Maybe it's better now. Skeptical about other tools because of junk code, accessibility not considered, design sprawl, etc.
59	1	Yes	The output changed based on the scope of the request and how many parameters we had to feed it on the fly. For example, asking it to create a whole page using our design system would result in really janky UI. But asking it to create chunks at a time and using specific tokens/density/best practices already defined in the system worked better. Have been thinking a lot about how to scale those chunks so that the context is the less probabilistic and AI only "colors within the lines."
60	3	Yes	private vs public scope
61	1	Yes	This experiment made it clear that AI is only as good as the decisions already encoded in the system. With a deliberately minimal, newly migrated design system, the AI filled undefined areas with choices that lacked design logic—from layout to hierarchy to typography. It reinforced that without explicitly defined constraints and patterns, AI doesn’t design, it guesses (badly).
62	2	Yes	It give great starting point, but we have not been able to explore too much, as we are struggling to get licences & dev capacity to dedicate to more explorations
63	2	No	I still think it might take more time to describe and correct whatever is generated than to do it myself. I might be wrong. But I'm glad i've stumbled into this discussion around AI and design systems.
64	0	I personally have not, but colleagues in other areas of my company have attempted it.	From what little visibility I was able to get into the work of my colleagues, the output was not great. It took many prompts, maxing out their usage limits, and still required further edits.
65	4	Yes	Got the look and feel but with some refinement as with most design work it felt good!
66	2	No	Current approved tooling at the time did not support leveraging a Figma Design System. There were other issues with security.
67	4	Yes	We're getting started with this and the early experiments are ripe with hallucinations. But we're also not surprised, as it demonstrates where we need to make improvements and iterate.
68	1	Not yet, but it's something we're planning for this quarter	2026 is the year we're getting licenses, tooling, and resourcing in order to dig into AI enablement
69	1	Yes	The amount of work necessary to map them hasn't been worth it so far
70	0	Yes	The output is not to the level I want it to be.
71	3	Yes	Our design system is meant to create consistent, predictable outcomes. When we’ve tried using AI, it often misses that goal. Instead of following the system’s rules, it tends to “guess” based on patterns from its training. The result usually looks ok, but it doesn’t truly match our system or intent. It feels more like an interpretation than a reliable application of the design system, which makes it hard to fully trust the output right now. We're working on the prompts... We'll get there. But I rather work on design that workign on the prompts.
72	5	Yes	It's taken a while to refine our outputs by optimizing the MCP server for our component library, but I'm pretty happy with it!
73	1	Yes	We need better and more robust documentation; also the system foundations itself are still pretty confused and patchy, so need rationalising. It's very annoying that Figma Make doesn't seem to really understand the Figma components in our library directly.
74	4	No	Design system maturity
75	2	Yes	Our Design team has mostly focused on prototyping with the design system components and tokens. We find there are many hallucinations in the output, some of which are the system's fault for not being properly set up to serve as context. It's cool that AI can fill the gaps with other knowledge it has of other systems, but that also causes quite a bit of confusion for people using it for prototyping without fully knowing the system. We need to rely on them to spot the inconsistencies created by the AI tooling rather than assuming that they're getting the complete system. There are other factors getting in the way of us fully realizing the promise of using the design system for building our UI with AI. Lack of time. Aside from just trying to understand how to structure MCP or build a skill for Claude, the time it takes to figure all of this out while trying to maintain and continue building a design system feels staggering. We've always been flying the plane and building it at the same time. Now we're being told we need to make it easy for everyone to get on the plane while it's 10,000 feet in the air.
76	2	Yes	It's okay. Nothing that would be able to be used for production, but gets quick ideas across for collaborative convos. Have significantly limited access to approved software to be able to use for this work.
77	2	Literally what I am starting this week.	Let you know in a week :)
78	4	Yes	It's ceiling is only as good as what the humans put in.
79	3	Yes	I feel optimistic about it. With the number of UIs at my organization, there is a lot we aren't able to do because we have a small team. But if we build things in the right way and leverage AI technology in the right way, we can use it to our advantage to audit products to see what and how the DS is being used. We hope to be able to identify patterns and use cases that have not been brought to light. And we hope to partner with product teams to refine our AI offerings to best meet their needs when it comes to AI. We have experimented a bit with generating AI creating UI and the results are okay but not great. There is still some odd things being created. So we know there is a lot of work to be done. AI can't create a UI without a really solid context so we working on the data sources and structure that need to be in place to produce accurate results.
80	1	No	Feel like it would end up being a lot of work in the long run to get the UI to standard
81	4	Yes	With the right pipeline and Figma code connect I can trust AI to only use existing hand crafted highly polished design system components which pushes the quality to create full screen UI output.
82	3	Yes	It’s improving but still not good.
83	5	Yes	I'm feeling quite confident, but the reason might be surprising: Most of the architecture that we've created in the past weeks is not AI-driven at all. We've been leveraging AI to get documentation in place where it was missing, we used it to write server scripts to do linting and fixing e.g. missing or wrong design tokens. We got the foundation so stable that I can now push from Figma and get clean React code, tokenized, published on Storybook... and with 100% accuracy (if designers miss a token, it will be auto-fixed). No MCP involved, by the way. We use a bridge plugin collecting various data as plain old JSON. That's pushed to our server, which executes a few powerful scripts. No AI in this step needed. The AI layer just consumes and patches up screens. The code can be imported by e.g. Figma Make looking at our Git repo. W3C Design tokens come from an npm package that updates itself with every merge into the main branch of our token set. That fully enables prototypes and live frontends based on our exact components and our exact tokens. Pipeline solved.
84	3	Yes	Our first design system is under way, so we’re literally only a few components in. But we’re building it with AI usage as a major priority. I’m currently designing this as my first official system (!!) and in the throes of figuring out how to write the instructions and guidelines for humans AND for AI 😅
85	2	Yes	I have been using it for suggestions based on the documentation and very rough in the early stages - using Claude. We also made an "AI Remix" tool inside Storybook which was very interesting but it was for a "hackathon" so not officially in use. Please note I cannot attend so will have to listen to the recording.
86	0	No	drowning in the most basic tasks, alignment, and work that were not enabled to move onto AI. trying to figure out how to solve this though!
87	3	No	The main blocker is not technical but epistemological: the design system is not formalized with the semantic precision required to function as a reliable operational context for AI. There are tokens, components, and documented rules, but there is a lack of explicit invariants, clear hierarchies, and measurable validity criteria. In that state, UI generation would be difficult to audit, reproduce, and correct, introducing more ambiguity than it removes. Where I do see value is in using AI in a deliberately constrained way for exploratory proposals outside of production, for example in CRO and rapid experimentation contexts focused on research. In that space, variability is an advantage rather than a risk, as long as exploration is not confused with production-ready output.
88	2	Yes	We have connected Builder.io to our design system and it's been pretty good for helping designers create realistic prototypes to test concepts with users. As for other approaches, our devs are exploring a Storybook MCP I am still not clear where the value lies in MCP servers for feature designers specifically? And what the optimal ways are for us to leverage AI powers to optimize the design workflows (other than AI prototyping tools).
89	3	I (as a designer) personally haven't had the chance to try it yet. But my dev partners (in my prev. org) have been experimenting with it for some time.	Seems to working mostly well, as long as designers ship with the exact library components in Figma. Doesn't always work well with deviation/non-library components.
90	3	No	na
91	3	Yes	We have tried using VS Code, Cursor, and Figma Make about our our DS, Everyhting seems to be working fine, but not upto mark. We are integrating Supernova MCP for documentation for better results, not tested yet.
92	2	Yes	The AI did well to use the design system. The design system was not ready for AI.
93	5	Yes	AI is good at pattern matching and repetition, so anything where it has to just apply design system patterns a whole lot tends to go OK.
94	2	No	Needed to much programming or code knowledge to know how to ai vibe-coding
95	2	Yes	We haven’t adopted AI-driven UI workflows yet because our reliance on PrimeNG and established design patterns makes it difficult to shift toward token-based or AI-generated customization
96	2	Yes	Feeling awful about the output of AI when leveraging our design system.
97	3	Yes	I’ve found that AI can generate code based UI much better than it can generate a mock-up. Which is disappointing because it can’t meet designers where they are, but it’s becoming the driving force to use new tools and workflows.
98	0	No	We have a lot of foundational work to do before we ca reliably connect to AI
99	1	No	We’re still in the foundational phase of setting up our system and the metadata that will eventually support AI use. While we’re excited about the potential to incorporate AI, our foundation isn’t in a strong enough state yet to support it effectively.
100	2	Yes	It does alright assembling small combinations of components but it has trouble with layout and larger pages.