Episode 060 Raw Data

	A	B
1	Thanks for your interest in The Question! Sign up to participate in future questions here. This file is the raw data from Episode 060, deep dive hosted on August 21, 2025 with Ben Callahan and Nathan Curtis. There are 95 answers. The question this week was: ----- Hello, system thinkers! I sat down with Nathan Curtis last week to find out what he’s been working on lately. Our conversation was a little winding, but we quickly settled into discussing AI, the future of our roles, and what we called the spectrum of precision. That last point focused on what it means to shift our mindsets from one where the output computers generate for us is deterministic to one where that same output is stochastic. We agreed that design system consumers must trust what we produce, so our processes must yield outputs that are deliberate, detailed, and stable over time. With AI, output is a prediction, not a rule. We ask it the same questions and get different answers given even subtle changes in context or moment. Nathan and I were considering where the line is for our work creating design systems. How much imprecision is too much. And, for a design system program where the goal is often stable, high quality, durable output, how much instability, lack of quality, or impermanence can we tolerate in our tools? Two weeks ago, we took your pulse on AI in design systems and you told us you were cautiously optimistic. This week, Nathan and I would like to better understand where AI is too imprecise and how that constrains its value to you. Welcome to Episode 060 of The Question. Is there an expectation in your organization that digital interface production will be done with AI in the near future? Where, in your daily work, is AI too imprecise for your needs—where is it failing you?
2	Question 1: Is there an expectation in your organization that digital interface production will be done with AI in the near future?	Question 2: Where, in your daily work, is AI too imprecise for your needs—where is it failing you?
3	Yes	Even with an MCP server accessing and retrieving production ready assets and knowledge, it's not accurate or trust worthy enough to embed into production. It's still only good enough for prototyping ideas with production ready components, it's not good enough to ship. The question is, should we ever allow AI to ship products into the hands of our customers? Should and will humans always be required to QA production? Even if we have sophisticated and automated design linting.
4	Yes	In code assistance, it can quite often get things almost right but not exactly. This can lead to unexpected side-effects on API or behaviour of design system components, which could well affect consumers and erode trust over time.
5	Yes	We work in a complicated enterprise product and the ai tools we have now only really build web pages it seems. That or we have to get better at prompting.
6	Yes	it is great for research and documentation, but I didn't find a good way to automate my work in figma itself. Also, when vibe coding, the design precision and love for details fails to my taste.
7	Yes	Pretty much everything. I'm a typical Design System designer who gives support to product teams, design components, write documentation etc. For the type of work I do other than summarising meeting notes and helping in the odd copy review I haven't found a use for AI that impactfully makes my work output better or faster.
8	Yes	Fails in thinking about the opportunities to really enhance the customer experience in context of their needs and how they intersect with business goals. Recommendations are good from a best practice standpoint but don't really bring something new to the design process - in fact what I've experienced using ChatGPT, Claude, Gemini and Figma AI has felt more like a junior designer and reminds me of the last several years managing designers. There is value in that - for one thing I love collaborating with designers and mentoring jr and mid-level. and Ive gotten interesting and helpful UI ideas from AI and have a productive experience refining those ideas as I would with JR/Mids, which brings me back to question one ... working with AI on UIs (not the code yet but let's check back in a few weeks) - when it's on track - has been similar in experience and quality as working with my junior and mid contractors in terms of getting the output I needed - but it was not as an enjoyable experience. If I were in my previous role of supervising designers I feel fairly confident my team would be shrinking in favor of AI tools. Hope this helps, I'll do my best to make the meeting!
9	No	I work in both product design and front-end development, and in my experience, there are currently almost no beneficial applications of AI in UI design. Figma Make does not adhere to our design system, nor does it consider the guidelines or use the components, even when prompted to do so. However, AI has improved our workflow tremendously on the code side. This is especially true when it comes to finding inconsistencies in the code and refactoring our CSS from Tailwind to CSS Modules.
10	Yes	weve been using AI to integrate our design system and drive adoption. the imprecision comes from the random way our current state is set up at times though. AI searches for class names and makes a call on what to replace where. It misses a lot of things or replaces components that dont make sense - but the problem isnt that the AI is wrong, its that it couldnt possible know better given the ruleset for our current state is...flawed at best. Its not the worst thing when its wrong, since it still does speed up our workflows, but the narrative that we can just let AI do it all is false because of its inaccuracies. Its not yet replacing full tasks in this case, but it is helping to automate the easy work and speed our workflows up outside of that, we havent been using AI a TON but my team did a hackathon where we ideated on a way to use it for our UAT tasks. this would involve training an image model where AI could understand more what components are needed in what areas by scanning the UI more visually. We havent gotten a chance to really dig our heels into this but this is the place i will be curious to see how accurate it can get.
11	Yes	I would say imprecise = failing only if we expect it to be precise. The more you work with it and understand where and how the limits are, you know how you need to manage it. No one trusts it to write shippable code right now, definitely not without proper review and testing. But it does offer a quick starting point, and can automate some repetitive tasks.
12	Yes	Generating front end screens with Figma MCP introduces variance, when we need strict adherence to our layout classes.
13	No	gpt-5 has made a nice shift, but not as earth shattering as the hype made it out to be. What I'm seeing is our product designers creating prototypes to get in front of clients, which is great, but they have glaring UX issues that are skewing the test results because they are incomplete. I also haven't been able to get a decent Figma → Code system working, yet.
14	Yes	At the moment AI isn’t too imprecise. This is largely because I’ve spent time experimenting on my own with AI tools rather than being in a team environment. That said my observations are that tools like Replit and Lovable give the visual illusion of being able to create somewhat complete or near finished outputs. At least to the untrained eye. Mostly these tools visually appear to reference the same toolkit of either material design or Schadcn. Even when I specifically reference a different design system toolkit - such as IBM Carbon Design System- layout mimic a generic box inside of box layout. Although it may use code from a library I specify it doesn’t seem to embrace the spirit of the design system. I start a new job next week and will be exploring the opportunity to build on what I’ve learned. I’ll let you know how I get on.
15	tbd	using customized data not generalization
16	No	For me, every AI tool I have encountered and tested has been massively underwhelming. I have occasionally found utility in brainstorming EXTREMELY rough copy drafts with ChatGPT, but even in that context, it's an additional step rather than a time savings. The information it generates is rarely accurate and therefore not trustworthy, and requires extensive fact checking and revisions before it can be used. All of the AI design tools I've tried have been abysmal. So much time wasted writing precise prompts and then refining those prompts over and over to still not generate anything usable. I always end up scrapping what's provided and creating what I need manually—usually much faster and definitely with better results.
17	No	Too non-deterministic still, and for foreseeable future for enterprise software.
18	Yes	There is nothing yet that can reliably generate editable designs (not code!) consistently from components. Even feeding components/tokens to no code platforms results in inconsistent and imprecise output. I would love to see AI be able to take on some of the heavy lifting of the more mundane tasks, e.g. token/variable creation, application and so on. Outside of general assistant type stuff on the periphery, it doesn't feel like AI can reliably sit within the actual production workflow for design or design systems within an established product. It's pretty great for early ideation and play but less so BAU or ongoing production work. Still exploring though!
19	Yes	Failing is too harsh. Appropriate expectations and learning to guide the tools has helped to lower “failure” rates.
20	No	We are getting v0 accounts soon, but the current plan is for it to only help with brainstorming and ideation. It isn’t currently equipped to follow our design system, which is limiting the usefulness of its output. And it doesn’t output in Figma, so we’ll still need to build the Figma mocks from scratch.
21	Yes	AI doesn't understand visual/white space very well at all in my experience.
22	No	We are not using AI yet
23	Yes, and no. We are considering training AI with our design system, UI patterns, and page templates. For teams without a dedicated design resource, this could help them ship higher-quality products faster.	AI assistance (e.g., a chatbot) will be too imprecise if it is generated from external sources rather than learning from our design system. At that point, we might as well tell our users to “Google it.”
24	There's currently a moratorium on AI use for the whole enterprise, so we haven't heard anything on how we could/should use AI.	N/A
25	Yes	Prototyping with our design system is hit and miss. This is important to our team to avoid vibecoding resulting in deviations from our system ending up in product. AI is good for discovery and exploration and communicating ideas but not always for precision. The time it takes to get things JUST right is not worth it.
26	Yes	It produces questionable results when generating new components. For example, it recently generated a component that had a dual API, which would've been hard for us to maintain. It also used non-ideal typescript practices (typecasing, any, etc).
27	Yes	Yes. We’re working on providing the right context layers (MCP, etc) to the AI tooling, but it feels like we’re still a ways off from it feeling like a slam dunk. Internally we have a lot of discussions about how we’re getting output from a super powerful (not to mention energy-intensive) tool that then requires a lot of clean up to get to something that is shippable. This is all to mention that we still have a lot of work to do internally. We’re currently learning from what other teams who are further along in their AI journey are putting into place and what they’re learning.
28		I have only been using it for experimentation and the output still needs a bit of fine tuning. When using it to help me code, it gives way more than I've asked for and as someone newer to development I don't have the knowledge to understand if what it is giving me is trash or best practice so I leave it out.
29	Yes	Currently, translating UI decisions into usable code that utilized the standards and foundation of the DS.
30	Yes	When I've used AI to identify key insights from a body of data, like a survey, for example, I have to be very careful that it is drawing specifically from the source provided, and not padding the results with generic and common insights that sound plausible at first glance. More than "imprecise" though, I have found AI generated code or documentation tend to be more expansive and verbose than needed. This has an impact on maintainability of code and dilutes the value of documentation unless effort is made to "sculpt" the result.
31	No	Using RAG LLM for a design system chatbot. Even good prompts will yield inconsistent (and often incorrect) responses. Users of the chatbot may not write good prompts.
32	Yes	Mapping of existing tokens and components in Figma to the tokens and components in codebases
33	Yes	Last year, AI burned me on my first earnest attempt to use it for a real task--something rudimentary that wouldn't have been a good use of my time. I was conducting a big color audit and refinement, and attempted to use LLM's to quickly categorize and arrange all my found HEX values. First, by arranging them into categories (blues, reds, purples, etc). Then, arranging from lightest to darkest within each category. The LLM's kept failing in various ways--putting some colors in the wrong category and/or arranging them in the wrong lightness order. When I dug deeper to figure out why, I discovered that the LLM's had wrong fundamental values (HSL, HSB, RGB, etc) for the HEX values I'd found. I wasted about an hour trying to get the LLM's to do something I could've just spend 20 minutes doing by hand.
34	Yes	The confidence with which it upholds bias. I do not trust it to not bake in all the 'isms' (like automatically and incorrectly gendering our CEO as a man... purely based on the CEO title).
35	Yes	It is getting better bit by bit, but generating codified views from figma design and with design system context is not quite there yet. We have been testing Figma make, Copilot and builder.io fusion. At times with few iterations however its getting quite close to usable code that correctly uses design system library.
36	Yes	Primarily in the ineffable "does it look like our design language" aspects. Code generation has gotten quite strong, but just like you can make a really ugly, inaccessible, badly designed UI with a design system, the LLMs don't really have context on design language or the subtle ways we put elements together to "look like" our brand.
37	Yes	It's slow. In the time and number of prompts you need to give an AI and then go back and forth, you probably could have created a usable output for [fill in the blank use case].
38	Yes	I work in healthcare, so the potential leaks of CBI is always a complicated--and sometimes nonstarter--detail (so, the precision on how/where/when any data put INTO an AI model will be used)
39	Sort of ... there's an assumption that aspects of our work, especially engineering work or, perhaps, production work within design system work, might at least be aided by AI more in the somewhat near future.	Even attempts at using AI for things like as a research assistant have been a bit underwhelming. For example, we've tried having AI analyze survey results and provide themes, etc. when we have questions with free form responses. What we found is that, at times, when testing with multiple AI's we get differing answers to even things like counts on the use of a given term within responses, and in some cases none of the AI answers were actually correct when we then re-checked the work. It's made us limit it's use to either low importance questions (where it's ok as long as it's close) or as a sounding board for comparative analysis for a strategist or designer working on their own.
40	Yes	Mainly naming things like components, properties or tokens (I dont have access to the magic of MCPs yet). But in fact, everything, when you don't give enough context and goal, but especially when you think that AI answer is task ready without any tweak from yourself to be perfect
41	Yes	Pixel perfection, for sure. Also, very precise instructions do not often convert to very precise outputs, no matter how precise you are. Seems to be a threshold there somewhere. If I'm in the need of something too detailed or specific, I know that I'll be kinda wasting my time if I go down the rails and be overly specific. It is easier for me to provide context and some pointers, then take AI's output and fill in the gaps myself. Also, I feel a bit less dumb by doing it that way. Regarding Design Systems and code implementation, I found that there's always a little bit of hallucination as you iterate; and if you keep on iterating and adjusting, it doesn't really get any better. There's a balance between how precise your prompts could be and the effort in adjusting such prompts towards a very precise result, versus just go and edit that particular line yourself, and call it a day.
42	No	I regularly use when creating UIs using design system components, but to create the components in the first place often requires more prompting and re-prompting than it's worth. The quality of output for brand new code just isn't up to the standard we expect for DS components. In design AI can be great for brainstorming, but even given huge knowledge base, it fails to understand the full context. It can create a vanilla name input in brand colors easily, but ask it to create something product specific and it almost always misses the mark.
43	It's mentioned but no expectation. We are more enthusiastic about the idea.	When it comes to the details.
44	Yes	I do not use AI in any way where it would not be precise enough for my needs. What does AI fail look like?
45	Yes	Understanding our design system.
46	Yes	It needs to generate scalable, consistent code based on design system variables. I know that people are starting to figure this out with MCPs and multi-layer architecture.
47	No	We’re only allowed to use Gemini at our company, so I would say lack of access to AI tools is a big blocker to even understanding the precision of the tools. As for Gemini, I’ve used it to draft ideas/content, synthesize research into insights documents, and other random tasks. The output has been fairly precise but definitely requires a good amount of review/critique and adjusting of prompts to get a quality output. The more context that can be given, the better the output.
48	Yes	I haven't used AI enough in my daily design system work for me to say where it's imprecise and failing me. I do wonder if as we start using AI more in our day to day work to replace some of our manual functions, will we then be supporting how the AI works to get a user results versus how the system works?
49	Yes	Early in the process: currently testing web -> native asset translation with automation signaling when native assets are best used
50	Yes	It can make design decisions and produce parts of a system, but doesn’t understand the human context especially in a complex org where we need to build alignment with the DS and use decisions to gain adoption and participation.
51	Yes	Using Cursor + Figma MCP server to code components. The QA required to clean the code is not less than the time it saves making it.
52	Yes	The area that we struggle with the most is the fact that you can give AI the same prompt multiple times, and get a different result. This is very clear when you have it generate UIs, you will often get completely different designs.
53	Yes	Generally it’s been a net positive. I’m in UX Program Management.
54	Yes	We're still early in our design system maturity journey (migrated a Sketch UI kit to Figma token and component libraries, but still need to update React components). A designer's recent attempt at using Figma Make for a quick prototype pulled in a random design system's component library to build our screens, which helped in terms of demonstrating the expected functionality, but wasn't actually usable as a prototype.
55	Some people could be looking at that and no one has mentioned it to me.	I've gotten frustrated trying to get AI to generate a precise visuals for me. When I ask for a graph and its not quite right, it doesn't know how to fix it when I ask it to revise. When I give it an image sample and say please make this for me but in a vectored format so I can edit it, I get something close but not exact.
56	No	Length of attention span, simultaneous big picture and small picture focus, context drift. Reliability.
57	Yes	In most areas I focus on use cases where AI focuses on repetitive low risks tasks. In places where there is creative thinking or complex coded outputs it feels the time required to provide safeguards, context and quality assurance. Area we had hoped to use it next were prototyping with styles and components that aligned with our DS, but the gaps it presented lowered confidence or required too much additional manual cleanup to be sustainable or provide significant value.
58	Yes	Garbage in, garbage out – the design system is not AI-ready, so it cannot provide accurate context for AI tools
59	No	Using Claude Code along with web apps production. Fails because it add too much technical debt. Fails to understand context precisely. Fails to provide practical and polished solutions out of the box. Still needs refinement. Context is crucial and has to be done properly. We need to get used to create proper prompts. Not sure if everyone will be up to that task. In the design systems world processes are key, introducing AI con be proven a time saver in certain aspects of you use RAG to train the AI to spit something like a component, but initial design, where it all starts is pretty much in diapers, and there is where I think AI will fail for now. Introducing changes processes trough prompting will be key, the question is how and we fail on that as well. We have experimented with this through variants, brands, modes and other areas with poor results.
60	Yes	It fails building my components, using the correct design tokens supplied by a repository, with accuracy. We only achieved about 80% accuracy with different tools, but less than 100% doesn't make sense to us. Hallucinated token values, unassigned tokens and loss of context are blockers for us at the moment. The AI fails to "understand" the systems, which makes sense in terms of it guessing about things it hasn't been specifically trained on.
61	Yes	Right now the tools used at work (we only have access to a few specific ones) the output struggles to follow the design system visually and it ignores the guidelines. Ideally it would also consider both goal metrics and what components/copy/content performs better for our audience. There are other tools like builder.io that seem very promising but I have not had the chance to test it myself. Also I would distinguish ideation/prototyping and code to be shipped (for now), meaning tolerance for prototyping is high vs none for shipping.
62	Yes	When looking at the same file over and over, it tells you that it’s doing something and then doesn’t do it, it never replicates the same code twice and most often tries from scratch.
63	Not all of the work, but definitely parts of it.	Inherent ethical problems, negative societal impact, sustainability concerns, hallucination.
64	Yes	(1) Synthesizing notes, especially as it relates to highlighting key insights from research; (2) Ethical and transparent opt in/out and data collection explanation, which is less on the AI itself and more on the implementation
65	Yes	(3) Energy demand, including severe consequences to the environment, public health, utility inflation, and strain on utility infrastructure
66	Yes	Due to company security measurements, we cannot use most of the AI features. Also, there's no concern about using the design system properly when talking about code that is generated randomly.
67	Yes	Currently available AI tools are missing the mark on what would empower designers in a major way. Most AI tools for interface creation offer one of two options: 1) AI ideation, where I describe the problem and some rough ideas for solutions, and the AI generates possible interface solutions. 2) High-fidelity design / implementation, where I upload sketches or wireframes and the AI makes them into high-fidelity prototypes or finished frontends. I flat-out do not need 1). I already have a powerful machine that knows interfaces and user needs and generates ideas. I call that thing my brain. 2) suffers from imprecision in a way that I imagine will be impossible to fix until we literally have chips in our brains. If I want a blue button of a specific size and border radius and colour contrast, it takes me half a minute to draw it up in Figma precisely how I like it. Achieving exactly the same result with AI is cost-intensive, not only because the AI can get things wrong, but because I already know precisely what I want – it doesn’t. Some design systems designers may be fine with giving up control and allowing an AI to decide precisely what the button should look like, even if it picks a font-size they wouldn’t have gone for had they built it themselves. I don’t see the benefit in giving up that control.
68	Yes	AI doesn’t fulfill any needs for me. It takes me out of the work of design and development and pushes it out of my mind or downstream. Yes, coding LLMs can produce valuable code, but set against the times it is wrong or slow, the net result is negative.
69	Yes	Quality content generation is falling behind for some people in our organization. I would love to use AI to automate our dev annotations and speed up handoff iterations.
70	No	AI is unpredictable depending on the type of work needed. And in some cases unreliable. In design systems data is key to help make informed decisions. Until this is addressed it is difficult to use it more than desired.
71	Yes	It has failed me with an AI-assisted contribution to a build script. The build script broke in an unexpected way, which caused our team to be blocked. I've also noticed that AI-generated design system documentation sometimes contains errors, and I'm concerned that when the documentation is fed into an AI coding tool, those errors will be reproduced at scale. Catching these errors in reviews is difficult and time-intensive. The AI-generated documentation covers 52 components in detail, so reviewing all of that documentation isn't practical. While reviewing the build script change, I did catch and fix an AI-generated error - but as I focused on solving that problem, I missed a larger issue that caused the build failure.
72	Yes	Actual rendering of the components from our system. I've either not learned how to get it all to work or in part there are still security issues I have to work around to get AI solutions to generate designs and prototypes with our components. Yet I have seen a product manager show me something that looks about 90% accurate and they think it is perfect. My trained eye quickly spots the differences but the average person that doesn't know the precise rules may not catch it.
73	No	Nothing comes to mind at the moment.
74	Yes	It’s been a net positive. I’m a UX Program Manager and it’s been useful.
75	Yes	We’ve only experimented with Figma make and the results were mediocre
76	No	not using AI for design work currently I use Microsoft CoPilot for some administrative task but no where else AI is a HOT topic in my compony so I'm attending to get Nathan's take
77	No	With larger code requests - occasionally Copilot will hallucinate answers to larger change requests. It will end up changing a whole bunch of stuff with the result being code that doesn't work. I've learned that larger, specific prompts, including file names, functions, and caveating everything with a note to propose changes prior to making actual changes in files helps.
78		No experience at the moment
79	No	Iterations. Like, reliability and consistency between sessions. Picking something up at later date to keep building on it doesn’t really work. Anything that takes time starts to suffer. AI loses context, gets gradually dumber, and needs more hand-holding, when you’d expect the opposite: more time and focus on something should lead to better output.
80	No	- Making learning MORE complicated - When I ask AI to help me figure out how to do a complex text task, ti often makes up things that aren't real. After giving this a try several times, I stopped resorting to ai to help with learning software. I was making the task take longer.
81	Yes	We're still in the middle of figuring out how to integrate AI. It's not so much AI failing us; it's more of us as a company trying to wrangle various teams and organizations working on AI to head in the same direction together. We have a lot of competing priorities in the AI space. In the areas where we have extensively used AI— AI has failed us because many things are hidden behind a "magic button." You enter prompt, click the button, and the magic happens. I wish AI tools gave us the ability to fine tune tasks and workflows. Figma Make is one of the best in class examples of this because you pinpoint a very specific area to work on. Everything shouldn't be a chatbot but most AI experiences and tools we use such as Co-pilot and Gemini and chat bot interfaces.
82	No	inaccurate summaries, impersoanl and recognisible tone of voice
83	Yes	I can't trust AI fully because it still makes things up, but I use it as a starting point and revise/customize as needed
84	No	AI is great as a search engine and highlighting best practices. It also allows me to figure out techniques and names of things I don't know about. But it definitely requires checking its work and reading the actual specifications---the hallucinations are real.
85	Yes	AI fails in the areas of nuanced precision. But this challenge is not exclusive to AI, is a human endeavor that requires skilled communication and facilitation to land the expected solution. So on the downside for AI, AI is still not above and beyond human capabilities in that space, and on the positive side for humans/design system designers.
86	Yes	It fails with troubleshooting code implementations and optimizations. Oftentimes it becomes a circular dance of repeating incorrect suggestions half an hour after stating their recommendation is not well structured. LLM depending, it can become a serious time drain.
87	Yes	As a content designer, I have not yet introduced an AI writing assistant into our content design system, in part because I worry it will spit out text that follows the rules but still isn't right. I guesss it's the fear of imprecise language that's holding me back.
88	No	AI struggles in areas where deep business context heavily influences design decisions. In our workflow, the design system is the backbone of everything we build, but most off-the-shelf AI tools aren’t aware of it and don’t integrate meaningfully with our established patterns or constraints. Without internal buy-in to engineer custom, integrated AI solutions, current tooling often feels disconnected from the realities of our product ecosystem and decision-making process.
89	Yes	I haven't been successful and using the Figma MCP server and producing anything of value.
90	Yes	I wear many hats in my organization. AI is failing me IT administration of our Office 365 environment. It is just too difficult for Microsoft to update the documentation and thus AI fails.
91	Yes	(Already answered, forgot to sign up for deep dive)
92	No	I don't think we've actually had AI do anything 100% from start to finish. It's helpful, but it's never been able to take a complex task from start to finish. Yes, it can help give me ideas for design tokens, but it doesn't know the whole picture and thus the names don't work everywhere. Yes, it can give us usage guidelines for our documentation, however, it doesn't know all the restrictions for the 19 platforms we use and all the one-offs. I view it as a single-task helper, not much bigger right now.
93	Yes	AI tools like ChatGPT and Claude Code have become indispensable in my day tod ay engineering work. I use them to plan how I'll tackle upcoming tasks, quickly bounce around ideas and polish documentation or rewrite messages. That said, where these tools fall short for me is in precision and reliability when working with technical codebases and with custom design systems. For example: * Sometimes they produce code that complies but doesn't align with our codebase best practices or my project's architecture. * They can overloop edge cases or give outdated answers * When working with Figma MCP and our custom design systems someitmes it will fallback to writing it's own components from scratch.
94	Yes	Accurately generating interface based on color tokens and other pre-defined specs. It also just likes to make up a bunch of stuff.
95	Yes	Design tokens in general (strategy, architecture, infrastructure, …)
96	Yes	When iterating, AI tends to lose the context it had built upon and starts suggesting solutions which are not always in proximity with the previous iteration. Having short conversations is helpful but at some point, i am required to take manual control when I’m building on top of an original idea, whether its code or design. In design systems, giving it context always help through component library access, code connect etc. but I have felt that the initial setup around building out the components, stabilising it etc is something that needs to be done manually.
97	No	In my daily work AI can sometimes be imprecise when the context I provide is too limited or not clearly defined, small ambiguities can lead to outputs that miss the mark. But even more than imprecision, I notice another challenge: when I rely on AI too heavily, my own critical thinking on complex topics can stagnate. The answers may be serviceable, yet the process of wrestling with nuance and forming my own perspective risks being overshadowed.
98
99
100